A personalized medication recommendation method, device, and storage medium based on multi-task learning

Abstract

The present invention discloses a personalized medication recommendation method, device and storage medium based on multi-task learning. The method includes the following steps: Step 1) Obtain the medical and health data of the user; Step 2) Obtain the medical and health feature representation of the user according to the medical and health data of the user; Step 3) Construct a medical knowledge graph, and obtain the drug feature representation according to the information of the drug entities in the knowledge graph; Step 4) According to the medical and health feature representation of the user and the drug feature representation, use the prediction function model to learn the interaction between the user's medical and health conditions and the drugs, and calculate the probability of interaction between the two; Step 5) Sort the drugs according to the probability obtained in Step 4), and recommend the top-k drugs to the user. Based on the idea of multi-task learning, the present invention fully integrates the effective information in the medical and health data and the medical knowledge graph, significantly improving the accuracy and interpretability of the personalized medication recommendation results.

Description

Technical Field

[0001] The present invention relates to artificial intelligence technology, and particularly to a personalized medication recommendation method, device, and storage medium based on multi-task learning. Background Art

[0002] In traditional medical practices, doctors often provide medications to patients based on their own experience and the patient's medical records. When different doctors judge the patient's physical condition, they are inevitably affected by personal subjective factors; moreover, there are significant individual differences in factors such as each patient's physical characteristics, physiological conditions, disease history, and genetic inheritance. Therefore, even for the same treatment plan and medication advice proposed by doctors for the same condition, different patients may still have completely different treatment effects or side effects. To address the existing problems, personalized drug recommendation has emerged, aiming to provide more accurate and scientific medication advice for patients.

[0003] In recent years, some drug recommendation methods based on deep learning have made recommendations by using historical medical events in the patient's medical records as auxiliary information, which has to some extent made up for the deficiencies of traditional drug recommendation methods. However, due to the diversity and complexity of medical and health data, such drug recommendation research has the problem of ignoring some important and effective information in the patient's health data. In addition, as the number of knowledge graphs in the medical field is increasing, there have also emerged many medication recommendation methods that use medical knowledge graphs as auxiliary features. However, these works still have problems such as insufficient utilization of effective information or introduction of noise in the utilization of graph information. Therefore, in the current context of the rapid development of the Internet, how to provide the most ideal drugs for patients through complex data information has become a major challenge faced by drug recommendation research. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a personalized medication recommendation method, device, and storage medium based on multi-task learning in view of the defects in the prior art.

[0005] The technical solution adopted by the present invention to solve its technical problems is as follows: A personalized medication recommendation method based on multi-task learning, comprising the following steps:

[0006] Step 1) Obtain the medical and health data of the user;

[0007] The medical and health data of the user includes physiological characteristic information such as the user's age, gender, and weight, the information of the diseases suffered, the information of the surgeries undergone, the historical medication information, and the physical examination information;

[0008] Step 2) Obtain the medical and health feature representation of the user according to the medical and health data of the user;

[0009] Step 3) Construct a medical knowledge graph, and obtain the drug feature representation according to the information of drug entities in the knowledge graph;

[0010] The medical knowledge graph is a triple of entity relationships of medical knowledge, where the entities include drugs, diseases, and adverse reactions; the relationships are the relationships between entities;

[0011] The medical knowledge graph is directly obtained from public data sets or professional medical databases, or indirectly constructed according to the drug instructions provided by the institution;

[0012] Step 4) According to the user's medical and health feature representation and the drug feature representation, use the prediction function to learn the interaction between the user's medical and health conditions and the drug, and calculate the probability of interaction between the two;

[0013] Step 5) Sort the drugs according to the prediction probability obtained in Step 4, and recommend the top-k drugs to the user.

[0014] According to the above solution, in the said Step 2), according to the user's medical and health data, obtain the user's medical and health feature representation, specifically as follows:

[0015] Based on multi-task learning of a deep neural network, construct a primary and secondary task learning network structure with shared parameters;

[0016] According to the user's medical and health data, set the prediction of whether the user's physical signs are abnormal as a secondary auxiliary task in the primary and secondary task learning network structure, and set the drug recommendation as the primary task in the primary and secondary task learning network structure. Use the feature analysis task to assist the drug recommendation task to obtain the enhanced user's medical and health feature representation.

[0017] According to the above solution, in the said Step 2), the primary and secondary task learning network structure includes: a bottom layer network and an upper layer network; at the bottom layer of the primary and secondary task learning network structure, extract features from the user's medical and health data, and the features include statistical features such as mean, standard deviation, and peak value, frequency domain features, time domain features, and time series features; after the feature extraction is completed at the bottom layer, at the upper layer of the primary and secondary task learning network structure, select a suitable classification model, and input all the extracted feature representations S(u i ) into the classification model, and finally obtain a prediction function f PA through learning and training, and calculate the probability that the patient's physical state is healthy: P u = f PA (S(u i )) According to the calculated prediction probability, obtain the user's medical and health feature representation.

[0018] According to the above solution, in the said Step 3), according to the information of drug entities in the knowledge graph, obtain the drug feature representation, specifically as follows:

[0019] Through the information complementary module, learn the potential interaction features between drug information and drug entity information, and then use the multi-hop knowledge graph embedding task to assist the drug recommendation task, making full use of the effective information in the medical knowledge graph to obtain a richer drug feature representation;

[0020] The drug information includes: the name of the drug, the usage method, the dosage, and whether there is an interaction between the drug and the user;

[0021] The drug entity information in the knowledge graph is the drugs related to the existing drugs in the knowledge graph, the applicable diseases, and the adverse reactions of the drugs;

[0022] According to the above solution, in step 3), according to the information of the drug entity in the knowledge graph, obtain the drug feature representation, specifically as follows:

[0023] Introduce an information complementary module that controls the cross-information transfer between drug recommendation and knowledge graph embedding. This module performs the following steps:

[0024] Cross step: Generate an interaction matrix between the drug's feature representation and the feature representation of the corresponding entity of the drug. The interaction matrix reveals the degree of interaction between the drug and the entity;

[0025] Compression step: Use the learned interaction matrix to map the embeddings of the drug and the entity to a unified feature space, that is, generate updated drug and entity embedding representations in the next layer;

[0026] For the feature vector v of the drug item and the entity vector e associated with item in the knowledge graph, first for their latent features and in the l-th layer, construct a d×d pairwise interaction: where is the cross-feature matrix for the l-th layer, and d is the dimension size of the hidden layer;

[0027] Then, by projecting this cross-feature matrix into the latent representation spaces of the drug and the entity, generate the feature vectors of the next layer: and where, and represent trainable weights, and are the corresponding bias vectors;

[0028] For drug v, use the information complementary module to encode the drug features of the user, and use the effective information in the medical knowledge graph to provide a rich feature representation for the drug, and encode it into a new feature vector v L。

[0029] According to the above solution, in step 3), multi-hop knowledge graph embedding is used to obtain deeper relationships of drug entities in the medical knowledge graph, and the prediction of the tail entity is completed;

[0030] First, use the information complementary module to encode the head entity h related to the drug v to obtain a rich representation of the drug entity Then use multiple non-linear layers to process the intermediate entity h n and the relationship r between each entity n to obtain the final head entity embedding representation h L and the relationship embedding representation r L , and finally combine the two and use a k-layer perceptron to predict the tail entity: is the vector representation of the finally predicted tail entity t. Finally, use a scoring function to calculate the score of the triple (h, r, t):

[0031] According to the above solution, the present invention also provides an electronic device, including: one or more processors; and a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors are caused to execute the method described in any one of the above solutions.

[0032] According to the above solution, the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the method described in any one of the above solutions is implemented.

[0033] The beneficial effects produced by the present invention are:

[0034] Based on the idea of multi-task learning, the present invention fully integrates the effective information in medical and health data and the medical knowledge graph, and significantly improves the accuracy of personalized medication recommendations and the interpretability of recommendation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0036] Figure 1 is the flowchart of the method of the embodiment of the present invention;

[0037] Figure 2 is the schematic diagram of the primary and secondary task network structure of the embodiment of the present invention;

[0038] Figure 3 is the schematic diagram of the principle of the information complementary module of the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the present invention will be further described in detail below in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0040] As Figure 1 shown, a personalized medication recommendation method based on multi-task learning includes the following steps:

[0041] Step 1) Obtain the medical and health data of the user;

[0042] The medical and health data of the user includes physiological characteristic information such as the user's age, gender, and weight, disease information, surgical information, historical medication information, and physical examination information;

[0043] Step 2) Obtain the user's medical and health feature representation according to the user's medical and health data;

[0044] Based on multi-task learning of a deep neural network, construct a primary and secondary task learning network structure with shared parameters;

[0045] According to the user's medical and health data, set the prediction of whether the user's physical signs are abnormal as a secondary auxiliary task in the primary and secondary task learning network structure, and set the medication recommendation as the primary task in the primary and secondary task learning network structure. Use the feature analysis task to assist the medication recommendation task to obtain an enhanced patient feature representation.

[0046] As Figure 2 , the primary and secondary task learning network structure includes: a bottom layer network and an upper layer network. On the bottom layer network, the input data is processed using shared parameters to ensure the effective transmission of information between different tasks; on the upper layer network, according to the different tasks, each task in this structure is trained using separate parameters to better adapt to their respective task requirements.

[0047] First, input the user's medical and health data into the bottom layer network, encode it using the shared parameters, then transfer the encoded output to the upper layer network, process it according to different task objectives, and finally use the result of predicting whether the user's physical signs are abnormal as one of the important inputs for the medication recommendation task to help obtain an enhanced user medical and health feature representation.

[0048] At the bottom layer of the primary and secondary task learning network structure, extract meaningful features from the user's medical and health data. The features include statistical features such as mean, standard deviation, and peak value, frequency domain features, time domain features, and time series features; after feature extraction is completed at the bottom layer, at the upper layer of the primary and secondary task learning network structure, select a suitable classification model, and represent all the extracted features S(ui ) Input it into the model, and finally obtain a prediction function through learning and training to calculate the probability that the patient's physical condition is healthy: P u = f PA (S(u i ))). According to the calculated prediction probability, obtain the encoded representation reflecting the patient's health status to accurately depict the patient's current physical condition.

[0049] Step 3) Construct a medical knowledge graph, and obtain the drug feature representation according to the information of drug entities in the knowledge graph;

[0050] The medical knowledge graph is a triple of entity relationships of medical knowledge, where the entities include drugs, diseases, and adverse reactions; the relationships are the relationships between entities;

[0051] The medical knowledge graph is directly obtained from public data sets or professional medical databases, or indirectly constructed according to the drug instructions provided by the institution;

[0052] Introduce an information complementary module to automatically learn the potential interaction features between drug information and drug entities, use the multi-hop knowledge graph embedding task to assist the drug recommendation task, make full use of the effective information in the medical knowledge graph, and obtain a richer drug feature representation. Drug information includes: the name of the drug, the usage method, the dosage, and whether there is an interaction between the drug and the user; the drug entity information in the knowledge graph is the drugs related to the existing drugs in the knowledge graph, the applicable diseases, and the adverse reactions of the drugs;

[0053] In the specific implementation, such as Figure 3 , introduce an information complementary module that automatically controls the cross-information transfer between drug recommendation and knowledge graph embedding.

[0054] This module includes two major steps: 1) The cross step, generate an interaction matrix between the drug's feature representation and the feature representation of its corresponding entity, and this interaction matrix reveals the degree of interaction between the drug and the entity; 2) The compression step, use the learned interaction matrix to map the embeddings of the drug and the entity to a unified feature space, that is, generate updated drug and entity embedding representations in the next layer, which is conducive to the joint promotion of the two tasks of drug recommendation and knowledge graph embedding.

[0055] For the feature vector v of drug item and the entity vector e associated with item in the knowledge graph, first construct a d×d pairwise interaction for their latent features and at the l-th layer: where is the cross - feature matrix for the l - th layer, and d is the dimension size of the hidden layer. Then, by projecting this cross - feature matrix into the latent representation spaces of drugs and entities, the feature vectors of the next layer are generated: and where and represent the trainable weights and bias vectors respectively. Among them, and represent the trainable weights, and are the corresponding bias vectors;

[0056] For the sake of convenience of description, the learning process in the information complementary module is represented as: [v l+1 ,e l+1 = ICU(v l ,e l ), and the suffixes [v] and [e] are used to distinguish the two outputs.

[0057] For drug v, the information complementary module is used to encode the user's drug features, making full use of the effective information in the medical knowledge graph to provide a rich feature representation for the drug, and encoding it into a new feature vector v L .

[0058] v L = E e～S(v) [ICU L (v,e)[v]], where S(v) represents the set of entities associated with drug v, and ICU(…) represents the information complementary module that enriches the feature representations of v and e related to the drug.

[0059] Utilize multi - hop knowledge graph embedding to obtain deeper relationships of drug entities in the medical knowledge graph and complete the prediction of the tail entity;

[0060] First, use the information complementary module to encode the head entity h v related to the drug to obtain a rich drug entity representation

[0061] where S(h) refers to the set of items associated with entity h, and ICU() is the information complementary module.

[0062] Then, use multiple non - linear layers to process the relationships r n between the intermediate entity h n and each entity to obtain the final head entity embedding representation h L and the relationship embedding representation r L , and finally combine the two and use a k - layer perceptron for the prediction of the tail entity: It is the vector representation of the tail entity t obtained from the final prediction. Finally, a scoring function is used to calculate the score of the triple (h, r, t):

[0063] The intermediate entities represent the one-hop entities, two-hop entities, and even n-hop entities expanded from the drug head entity. These entities can be entities such as diseases, adverse reactions, genes, and other drugs, and there are corresponding relationships connecting these entities.

[0064] Step 4) Input the user's medical and health feature representation and the drug feature representation into the drug prediction module, use the prediction function to learn the interaction between the user's medical and health conditions and the drug, and calculate the probability of their interaction;

[0065] After obtaining the final feature representations of the user and the drug, the prediction function can be used to learn the interaction between the two.

[0066] First, divide the processed data set into a training set and a test set. Obtain the final patient feature representation and drug feature representation according to the above method. Select a suitable prediction model to train these two types of feature representations. During this process, optimize the model through evaluation indicators such as accuracy, recall rate, and F1 value. Finally, use the trained model to calculate the patient feature representation and the drug feature representation to obtain the probability score of the interaction between the drug and the patient;

[0067] Calculate the probability of interaction between user u and drug v:

[0068] Common prediction model functions include inner product operation, dot product operation, concatenation operation, neural network, attention mechanism, etc. Through these operations, a more effective interaction between the user and the drug is learned, so as to recommend more suitable drugs to the user.

[0069] Step 5) Sort the drugs according to the prediction probability obtained in Step 4, and recommend the top-k drugs to the user.

[0070] The whole process of the present invention includes three key parts: First, obtain the user's medical and health feature representation. Through the given user's medical and health data, in the underlying network of the primary and secondary task learning network structure, extract meaningful features from the original health data and perform encoding representation, and combine the encoding representation indicating the user's health status obtained from the patient's physical sign analysis and the original feature vector u to finally obtain a comprehensive and integrated patient feature representation u LSecond, obtain the drug feature representation. For drug v, use the information complementary module to encode the user's drug features, make full use of the effective information in the medical knowledge graph, provide rich feature representations for the drugs, and encode them into a new feature vector v L Finally, make a prediction. After obtaining the final feature representations u L and v L of the patient u and the drug v, combine the two and use a prediction function to learn the interaction between them, and calculate the probability that the user u interacts with the drug v: where f RS represents the prediction function. The obtained prediction probability intuitively reflects the potential utility of the drug for the patient. Therefore, based on these probabilities, the drugs can be ranked, and the drugs with higher probabilities will be considered more suitable for the patient and will be given priority to be recommended to the patient.

[0071] Through the description of the above embodiments, those skilled in the art can clearly understand that the above embodiment methods can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases, the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a read-only memory / random access memory, magnetic disk, optical disc), and includes several instructions for causing a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present invention.

[0072] It should be understood that those of ordinary skill in the art can make improvements or transformations according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention.

Claims

1. A personalized medication recommendation method based on multi-task learning, characterized in that: The following steps are involved: Step 1) Obtain the user's medical health data; The user's medical health data includes the user's age, gender, and weight, including physiological characteristics, diseases, surgeries, historical medications, and physical examinations; Step 2) obtaining a user's medical health feature representation based on the user's medical health data; Step 3) construct a medical knowledge graph, and obtain drug feature representation based on the information of drug entities in the medical knowledge graph; The medical knowledge graph is an entity-relationship triple of medical knowledge, where entities include drugs, diseases, and adverse reactions; and relationships are relationships between entities. Among them, according to the information of drug entities in the knowledge graph, the drug feature representation is obtained, as follows: An information complementation module is introduced to control the cross-information transfer between medication recommendation and knowledge graph embedding, which performs the following steps: In the crossover step, an interaction matrix is ​​generated based on the feature representation of the drug and the feature representation of the entity corresponding to the drug. The interaction matrix reveals the degree of interaction between the drug and the entity. The compression step uses the learned interaction matrix to map the embeddings of drugs and entities to a unified feature space, i.e., generating updated embedding representations of drugs and entities in the next layer; For the feature vector v of the drug item and the entity vector e associated with the item in the knowledge graph, first, their potential features are and Construct d×d pairwise interactions: in is the cross feature matrix of the lth layer, and d is the dimension size of the hidden layer; Then, the feature vector of the next layer is generated by projecting this cross-feature matrix into the latent representation space of drugs and entities: and in, and represents the trainable weights, and is the corresponding deviation vector; For drug v, the information complementation module is used to encode the user's drug features, and the effective information in the medical knowledge graph is used to provide a rich feature representation for the drug, which is encoded into a new feature vector v L ; Step 4) Based on the user's medical health feature representation and the drug feature representation, the prediction function model is used to learn the interaction between the user's medical health and the drug, and the probability of the interaction between the two is calculated; Step 5) Sort the drugs according to the probabilities obtained in step 4) and recommend the top-k drugs to the user.

2. The personalized medication recommendation method based on multi-task learning according to claim 1, characterized in that: Medical knowledge graphs are directly obtained from public data sets or professional medical databases, or indirectly constructed based on drug instructions provided by institutions.

3. The personalized medication recommendation method based on multi-task learning according to claim 1, characterized in that: In step 2), the user's medical health characteristics are obtained based on the user's medical health data, as follows: Based on multi-task learning of deep neural networks, a parameter-sharing primary and secondary task learning network structure is constructed; According to the user's medical health data, predicting whether the user's physical signs are abnormal is set as a secondary auxiliary task in the primary-secondary task learning network structure, and medication recommendation is set as the main task in the primary-secondary task learning network structure. The feature analysis task is used to assist the medication recommendation task to obtain the enhanced user medical health feature representation.

4. The personalized medication recommendation method based on multi-task learning according to claim 3, characterized in that: In the step 2), the primary and secondary task learning network structure includes: a bottom layer network and an upper layer network; at the bottom layer of the primary and secondary task learning network structure, features are extracted from the user's medical health data, and the features include statistical features including average value, standard deviation, peak value, frequency domain features, time domain features and time series features; after the feature extraction is completed at the bottom layer, a suitable classification model is selected at the upper layer of the primary and secondary task learning network structure, and all the extracted features are represented as S(u i ) is input into the classification model, and a prediction function f is finally obtained through learning and training. PA , calculate the probability that the patient's physical condition is healthy: P u =f PA (S(u i )), obtain the user's medical health feature representation based on the calculated prediction probability.

5. An electronic device, characterized in that: include: one or more processors; as well as a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors are caused to execute the method according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images