Drug recommendation method and system based on multivariate information perception fusion

Abstract

The invention discloses a drug recommendation method and system based on multivariate information perception fusion, and belongs to the technical field of natural language processing, and the method comprises the steps: taking a health record of a patient as input, and obtaining vector representations of diagnosis, operation and drugs of the patient through a recurrent neural network; calculating a correlation score of historical access and current access, and distributing different weights to carry out representation construction; modeling drug molecular structures and drug substructures to obtain a related drug substructure list; performing modeling representation on the HER and the DDI graph by using GNN for controlling the DDI rate; and integrating the information to carry out drug recommendation prediction and calculating drug scores. The invention aims to overcome the limitation of the prior art, provides an innovative drug recommendation method, verifies the effectiveness of incorporating historical drug use knowledge into drug recommendation, and realizes more accurate drug recommendation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug recommendation, and in particular to a drug recommendation method and system based on multi-element information perception fusion. BACKGROUND

[0002] With the continuous development of deep learning technology, intelligent medical decision-making has shown encouraging results. The digital development of the medical field has also led to the accumulation of medical data such as electronic health records (EHR), laying a good foundation for the development of artificial intelligence in medical care.

[0003] Joint drug recommendation is to recommend personalized drug combinations to patients based on their electronic health records (EHR), which is an important task in medical care. The essence is to solve the problem of optimizing drug combinations with high precision under safety constraints. The complexity of the number of drugs, drug interactions, and recommended doses poses a serious information overload problem for doctors. In order to help doctors make wise decisions when recommending drugs to patients, researchers use deep learning technology to automatically recommend drugs based on EHR. Deep learning technology can effectively map various types of information into hidden vectors, and through heterogeneous information networks, embedding representation and other methods, auxiliary information can be integrated into the recommendation process, greatly improving the accuracy of the recommendation. Combining deep learning technology for drug recommendation can effectively mine the correlation between patient physiological characteristics and drug use.

[0004] Due to the strong professional nature of the drug recommendation task, it is necessary to integrate external medical domain knowledge to enhance the credibility of the recommendation results. At present, most drug recommendation models rely on data-driven methods, which use artificial intelligence algorithms to analyze patient data and generate personalized drug recommendations. Although these methods have made preliminary progress, they cannot explicitly model the complex relationships between drugs, diseases, and patient characteristics, resulting in poor explainability of the recommendations. In order to solve this problem, researchers have proposed various methods to incorporate domain knowledge into drug recommendation models. For example, develop a knowledge graph to represent the relationships between medical entities in a structured manner. The knowledge graph can be used to model various types of entity relationships, such as drug-drug interactions, drug-disease associations, and drug-gene interactions. Existing methods mostly focus on learning global temporal dependencies in patient EHR sequences, ignoring local temporal dependency information at the level of visits. However, in many cases, doctors often judge the progress of a patient's condition based on recent visit records, so the latter is equally important as the former. The above defects limit the representation learning ability of traditional drug recommendation models, resulting in poor performance of traditional drug recommendation models in drug recommendation tasks, and thus unable to recommend accurate and safe drug combinations for patients. SUMMARY

[0005] Therefore, the application provides a drug recommendation method and system based on multi-element information perception fusion, which can effectively and quickly recommend accurate and safe treatment drugs for patients.

[0006] To this end, the application provides the following technical solutions: A drug recommendation method based on multi-element information perception fusion comprises the following steps: The patient health record is taken as input to obtain vector representation of patient diagnosis, surgery and drug through a recurrent neural network; The correlation score of historical access and current access is calculated, and different weights are assigned for representation construction; The drug molecular structure and drug substructure are modeled to obtain a related drug substructure list; The GNN is used to model and represent the HER and DDI graph for controlling the DDI rate; The information is integrated for drug recommendation prediction and drug score calculation.

[0007] The further improvement of the technical scheme of the application is that the patient health record at least includes the diagnosis information, surgery information and historical drug information of the patient; the vector representation includes the following contents: For the tth visit of the patient, the tth diagnosis and surgery record of the patient and the t-1th drug record are selected, that is, ; Three different embedding tables are designed, , and , wherein d represents the embedding dimension, , and represent the number of diagnoses, surgeries and prescription drugs respectively; The patient's each access record is mapped into an embedding space to obtain respective embedding vectors: (1) (2) (3) The recurrent neural network is used to model the longitudinal medical record of the patient to obtain the output layer and the hidden layer of the recurrent neural network: (4) (5) Wherein, o and h represent the output layer and the hidden layer respectively; The splicing function is used to integrate the patient's diagnosis and surgery information to obtain: (6) R is represented for the patient.

[0008] The further improvement of the technical scheme of the present application is that the correlation score of the historical access and the current access is calculated, and different weights are assigned for representation construction, comprising: calculating the correlation relationship between the historical medical record of the patient and the current medical record, and modeling the historical drug representation of the patient and the patient representation through the correlation score; The records related to the diagnosis information are encoded by attention as: (7) Then, the combine combines the diagnosis vector and the diagnosis vector to obtain the final diagnosis vector: (8) p representing the final diagnosis vector; The embedding of each visit moves one position backward, so that each visit can be compared with all previous visits; the similarity score of each visit of the diagnosis and procedure embedding is calculated respectively by matrix multiplication; then, the relative score of each visit with all previous visits is obtained and returned; finally, the patient vector is obtained: (9) wherein, w 1 and w 2 represent weight parameter values; Using the history record is tailored for the object with previous cases, which helps the model to identify the historical information that is strongly related to the current recommendation; the historical information and the current information are represented as follows: (10) (11) A distance loss function is developed, which uses relative entropy to quantify the correlation between the current visit and the previous visit by measuring the divergence between their distributions: (12).

[0009] The further improvement of the technical scheme of the present application is that if the patient is the first visit, the correlation information is not calculated.

[0010] The further improvement of the technical scheme of the present application is that for the input drug, the representation learning of the drug molecular structure is performed by GNN; the GNN uses a neighbor aggregation strategy to iteratively update the node features, that is, the representation of the node is updated by aggregating the representations of the neighbors: (13) wherein, h representing the update; The representation of the drug molecule is generated by n layers of GNN; then, all the node features of the nth layer are aggregated by the readout function: (14) The global structure and molecular substructure are self-attended to obtain a set of relevant drug structures; the method of BRISC is used to analyze the drug molecule and retain the most critical structure; then, the patient representation is converted into a substructure vector through a linear transformation layer and an activation function to measure the importance of each substructure: (15) The drug structure embedding table is represented, which is a new drug structure embedding table that encodes the drug structure and obtains a drug representation s; at the same time, the diagnosis vector and the treatment vector of the patient are input, and the cross-score between visits is calculated through a pre-defined function; the diagnosis and treatment vectors are sent to the correlation calculation module; the attention of the visit is modeled as: (16) Wherein, represents a linear function.

[0011] The further improvement of the technical scheme of the application is that the graph convolutional neural network is used to encode the graph: (17) (18) (19) G EHR The vector encoding of the EHR graph is represented; G DDI The vector encoding of the DDI graph is represented; The graph loss is defined as: (20).

[0012] The further improvement of the technical scheme of the application is that the information is integrated for drug recommendation prediction and drug score calculation, including: The correlation score and the related drug substructure list obtained by the correlation calculation are fused by feature perception, and a feedforward neural network and An activation function is applied, and the final weight is as follows: (21) m t The probability of the occurrence of a certain drug in the prescription is represented, a multi-hot prediction vector is obtained, and the final drug recommendation result is obtained by selecting the items exceeding the predetermined threshold.

[0013] The further improvement of the technical scheme of the present application is that the threshold of the drug score is set to 0.5, and the drug is considered to be recommended if the score is greater than 0.5, otherwise the drug is not recommended.

[0014] A drug recommendation system based on multi-information perception fusion comprises: A patient representation module is configured to obtain vector representations of patient diagnoses, surgeries and drugs by a recurrent neural network taking patient health records as input. A correlation calculation module is configured to calculate a correlation score of historical visits and current visits, and assign different weights to represent construction. A drug attention module is configured to model drug molecular structures and drug substructures, and obtain a list of related drug substructures. A graph representation learning module is configured to model and represent HER and DDI graphs using GNN to control the DDI rate. A drug recommendation module is configured to integrate information for drug recommendation prediction and calculate drug scores.

[0015] Advantages and positive effects of the present application: 1. The present application performs deep mining on patient health record information, calculates the correlation weight between different access records and current records, and integrates the weight information in patient representation modeling and drug molecular embedding representation, thereby reducing the noise caused by irrelevant information by fusing high-correlation historical records.

[0016] 2. The present application proposes a multi-angle joint distance module comprising a correlation calculation module and a drug attention module, which extracts highly relevant historical information from the current access of EHR, and can extract patient and drug features from selected historical information.

[0017] 3. The present application designs a graph representation learning module to explore the internal correlation between drug molecules by learning from EHR graphs and DDI graphs.

[0018] 4. The present application introduces medical external knowledge and designs a drug interaction (DDI) loss function in the model training process, which greatly improves the accuracy and safety of drug recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments or prior art description. Obviously, the drawings described below are only some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0020] Figure 1 Flowchart of the present application; Figure 2 Model diagram of the present application. DETAILED DESCRIPTION

[0021] In order to make the person skilled in the art better understand the present application, the following will combine the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0022] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0023] As Figure 1 shown, a drug recommendation method based on multi-element information perception fusion includes the following steps: Step 1, input the patient's health record into the recurrent neural network to obtain the vector representation of the patient's diagnosis, surgery and drug; Specifically, the patient's health record includes the patient's diagnosis and surgery information, and the patient's historical drug information. The patient representation module obtains the patient's current diagnosis and surgery information and the last time's drug information from the MIMIC-III dataset through preprocessing, and models the vector representation of the patient's diagnosis, surgery and drug through recurrent neural network (RNN).

[0024] For the tth visit of a patient, the tth diagnosis and surgery records of the patient and the t-1th medication records are selected, i.e. Three different embedding tables are designed, , and , where d represents the embedding dimension, , and represent the diagnosis, surgery and number of prescription drugs, respectively. The embedding vectors of each visit record of the patient are obtained by mapping the patient's visit record into the embedding space: (1) (2) (3) The output layer and hidden layer of the recurrent neural network are obtained by modeling the longitudinal medical records of the patient using the recurrent neural network: (4) (5) where, o and h represent the output layer and the hidden layer, respectively; The diagnosis and surgery information of the patient is integrated using a concatenation function to obtain: (6) where, R represents the patient representation.

[0025] Step 2, calculate the correlation score of the historical visit and the current visit, and assign different weights to represent the construction; specifically including: The correlation between the historical visit record of the patient and the current visit is calculated, and the historical drug representation of the patient and the patient representation are modeled by the correlation score; The records related to the diagnosis information are encoded by attention as: (7) Then, the combine harvester combines the diagnosis vector and the diagnosis vector to obtain the final diagnosis vector: (8) where, p represents the final diagnosis vector; Each visit embedding moves one position backward, so that each visit can be compared with all previous visits; the similarity scores of each visit of the diagnosis and procedure embedding are calculated by matrix multiplication respectively; then, the relative scores of each visit with all previous visits are obtained and returned; finally, the patient vector is obtained: (9) wherein, w 1 and w 2 represent weight parameter values; The history record is tailored for the object with previous cases, helping the model to identify the historical information that is strongly related to the current recommendation; the historical information and the current information are represented as follows: (10) (11) A distance loss function is developed, using relative entropy to quantify the relevance between the current visit and the previous visits, by measuring the divergence between their distributions: (12) If the patient is visiting for the first time, the relevance information is not calculated; the calculation of relevance aims to eliminate the historical visit information irrelevant to the current visit, removing the noise brought by irrelevant information.

[0026] Step 3, modeling the drug molecular structure and drug substructure, obtaining a list of related drug substructures; Specifically, it includes: For the input drug, the GNN is used for the representation learning of the drug molecular structure; the GNN uses a neighbor aggregation strategy to iteratively update the node features, i.e., updating the representation of the node by aggregating the representations of the neighbors: (13) wherein, h represents the update; The representation of the drug molecule is generated through n layers of GNN; then, all the node features of the nth layer are aggregated through the readout function: (14) The global structure and the molecular substructure are self-attentive to obtain a set of related drug structures; the method of BRISC is used to analyze the drug molecule and retain the most critical structure; then, the patient representation is converted to a substructure vector through a linear transformation layer and an activation function to measure the importance of each substructure: (15) It is represented as a drug structure embedding table, which is a new drug structure embedding table that encodes the drug structure and obtains the drug representation s; at the same time, the diagnosis vector and the treatment vector of the patient visit are input, and the cross-score between visits is calculated through a pre-defined function; the diagnosis and treatment vectors are sent to the correlation calculation module; the visit attention is modeled as: (16) wherein, representing a linear function.

[0027] Step 4, using GNN to model the representation of HER and DDI graph for controlling DDI rate; Specifically includes: Two kinds of knowledge graphs are introduced, drug-drug interaction (DDI) graph and patient electronic health record (EHR) graph, where the DDI graph represents the adverse drug reactions between different drugs, and the EHR graph represents the drug coexistence relationship in the prescribed drugs.

[0028] Graph convolutional neural network (GNN) is used to encode the graph: (17) (18) (19) G EHR Vector encoding representing the EHR graph; G DDI Vector encoding representing the DDI graph; The graph loss is defined as: (20) Step 5, integrate information for drug recommendation prediction and calculate drug score; Specifically includes: Feature-aware fusion is performed on the correlation score and related drug substructure list calculated by correlation, and feedforward neural network and Activation function, the final weight is as follows: (21) m t The probability of the occurrence of a certain drug in the prescription is represented, obtaining a multi-hot prediction vector, and selecting items exceeding a predetermined threshold to obtain the final drug recommendation result.

[0029] The threshold of drug score is set to 0.5, and if the score is greater than 0.5, the drug is recommended, otherwise the drug is not recommended.

[0030] As Figure 2 shown, a drug recommendation system based on multi-element information perception fusion, that is, a drug recommendation model including patient representation module, correlation calculation module, drug attention module, graph representation learning module and drug recommendation module is built, specifically including: The patient representation module is used to input the patient health record into the recurrent neural network to obtain the vector representation of the patient's diagnosis, surgery and drugs; Specifically, the patient data in the MIMIC-III dataset is preprocessed to obtain the diagnosis and surgery information of the current visit of the patient and the drug information of the last visit, and vector representations of the patient diagnosis, surgery and drug are obtained through a recurrent neural network (RNN) modeling.

[0031] A correlation calculation module is configured to calculate a correlation score of the historical access and the current access, and assign different weights to represent the construction; A drug attention module is configured to model the drug molecular structure and the drug substructure, and obtain a related drug substructure list. Specifically, the correlation calculation module and the drug attention module constitute a multi-angle joint distance module. In this part, a single loss cannot accurately calculate the multi-factor influence in the whole process of drug recommendation. Especially for complex medical problems, the learning process of each step of the neural network needs to be controlled in detail, and multiple variables need to be adaptively fused to obtain more accurate, comprehensive and reliable results. The adaptive fusion of the drug attention submodule and the global distance network submodule helps drug recommendation.

[0032] A graph representation learning module is configured to use GNN to model and represent HER and DDI graphs to control the DDI rate. Specifically, two kinds of knowledge graphs are introduced, a drug-drug interaction (DDI) graph and a patient electronic health record (EHR) graph. The DDI graph represents the adverse drug reactions between different drugs, and the EHR graph represents the drug coexistence relationship in the prescribed drugs. The DDI graph and the EHR graph are modeled by a graph convolution network (GCN) to obtain drug interaction representation and patient electronic health record representation.

[0033] A drug recommendation module is configured to integrate information for drug recommendation prediction and calculate drug scores.

[0034] Specifically, the related patient embedding representation and drug information obtained by the correlation calculation module and the drug attention module are taken as input, and the drug recommendation module integrates the drug interaction information integrated by the graph representation learning module to make a prediction and obtain a drug prediction score. The final drug set is obtained by comparing the drug prediction score with a threshold.

[0035] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for part or all of the technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A drug recommendation method based on multivariate information perception fusion, characterized in that: The following steps are involved: The patient's health records are used as input through a recurrent neural network to obtain vector representations of the patient's diagnosis, surgery, and medication; Calculate the relevance scores of historical visits and current visits, and assign different weights for representation construction; Model the drug molecular structure and drug substructure to obtain a list of related drug substructures; Use GNN to model HER and DDI graphs to control the DDI rate; Integrate information to make drug recommendation predictions and calculate drug scores.

2. The drug recommendation method according to claim 1, characterized in that: Patient health records include at least the patient's diagnosis information, surgery information, and medication information from the patient's previous visits; the vector representation includes the following: For the patient's t-th visit, select the patient's t-th diagnosis and surgery record, and the t-1-th medication record, that is, ; Design three different embedding tables, , and , where d represents the embedding dimension, , and represent the number of diagnoses, surgeries, and prescription drugs, respectively; Map each patient's visit record into the embedding space to obtain their respective embedding vectors: （1） （2） （3） Use a recurrent neural network to model the patient's longitudinal medical history and obtain the output layer and hidden layer of the recurrent neural network: （4） （5） in, o and h Represent the output layer and hidden layer respectively; Use the concatenation function to combine patient diagnosis and surgery information to obtain: （6） R For patients.

3. The drug recommendation method according to claim 1, characterized in that: Calculate the correlation scores between historical visits and current visits, and assign different weights for representation construction, including: calculating the correlation relationship between the patient's historical medical records and current visits, and modeling the patient's historical drug representation and patient representation through the correlation score; Records related to diagnostic information were encoded via attention as: (7) The combiner then combines the diagnosis vector and the diagnosis vector to get the final diagnosis vector: (8) p represents the final diagnosis vector; Each visit embedding is shifted back one position so that each visit can be compared with all previous visits. The similarity score of each visit is calculated for the diagnosis and procedure embeddings separately through matrix multiplication. Then, the relative score of each visit to all previous visits is obtained and returned. Finally, the patient vector is obtained: (9) in, w 1 and w 2 represents the weight parameter value; The usage history is tailored for objects with previous cases, helping the model identify historical information that is strongly relevant to the current recommendation; historical information and current information are represented as follows: (10) (11) We develop a distance loss function that uses relative entropy to quantify the relevance between the current visit and previous visits by measuring the divergence between their distributions: (12)。 4. The drug recommendation method according to claim 3, characterized in that: If the patient was visiting the clinic for the first time, no relevance information was calculated.

5. The drug recommendation method according to claim 1, characterized in that: For the input drug, GNN is used to learn the representation of the drug molecular structure; GNN uses a neighbor aggregation strategy to iteratively update the node features, that is, to update the node representation by aggregating the representation of the neighbors: (13) in, h Represents update; The representation of the drug molecule is generated by an n-layer GNN; then, all node features of the n-th layer are aggregated through the readout function: (14) The global structure and molecular substructures are self-focused to obtain a set of relevant drug structures. The BRISC method is used to analyze the drug molecules and retain the most critical structures. Then, the patient representation is converted into a substructure vector through a linear transformation layer and an activation function to measure the importance of each substructure: (15) Represents a drug structure embedding table, which is a new drug structure embedding table that encodes the drug structure and obtains the drug representation 𝑠; at the same time, the diagnosis vector and treatment vector of the patient's visit are used as input to calculate the cross-score between visits through a pre-defined function; the diagnosis and treatment vectors are sent to the correlation calculation module; the visit attention is modeled as: (16) in, Represents a linear function.

6. The drug recommendation method according to claim 1, characterized in that: Encoding graphs using graph convolutional neural networks: (17) (18) (19) G EHR Vector encoding representing the EHR graph; G DDI Vector encoding representing the DDI graph; The graph loss is defined as: (20)。 7. The drug recommendation method according to claim 1, characterized in that: Integrate information to predict drug recommendations and calculate drug scores, including: The correlation scores obtained by correlation calculation and the list of related drug substructures are fused with feature perception, and a feedforward neural network and Activation function, the final weights are as follows: (21) m t The probability of a drug appearing in the prescription is expressed as a multi-hotspot prediction vector, and the items exceeding the predetermined threshold are selected to obtain the final drug recommendation result.

8. The drug recommendation method according to claim 7, characterized in that: The threshold of the drug score is set to 0.

5. If the score is greater than 0.5, the drug is recommended; otherwise, the drug is not recommended.

9. A drug recommendation system based on multi-information perception fusion, capable of implementing the drug recommendation method according to claims 1-8, characterized in that: include: A patient representation module, which uses patient health records as input and generates vector representations of patient diagnoses, procedures, and medications through a recurrent neural network. The correlation calculation module is used to calculate the correlation scores between historical access and current access, and assign different weights for representation construction; Drug attention module, used to model drug molecular structure and drug substructure, and obtain a list of related drug substructures; A graph representation learning module is used to model HER and DDI graphs using GNN to control the DDI rate; The drug recommendation module is used to integrate information to make drug recommendation predictions and calculate drug scores.

Images