Drug recommendation method, system and equipment based on depth matrix decomposition, and medium
By employing a drug recommendation method based on deep matrix factorization, combined with attention mechanisms and drug interaction thresholds, the problem of insufficient safety in drug recommendation models is addressed, resulting in safer and more accurate drug recommendations.
Patent Information
- Application Number
- CN202411235467.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-04
- Publication Date
- 2024-11-29
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing drug recommendation models have improved the accuracy of recommending drug types, but they neglect adverse interactions between drugs, resulting in insufficient safety in drug recommendations and failing to meet the safety and reliability requirements of drug recommendations.
We employ a deep matrix factorization-based approach, constructing a bilateral decomposition model and introducing an attention mechanism and drug interaction threshold to control the occurrence of adverse drug reactions. We also utilize multilayer neural networks and batch normalization techniques to enhance the model's expressive and fitting capabilities.
It improves the safety and practicality of drug recommendations, enables more accurate recommendations of multiple drugs, reduces the probability of adverse drug reactions, and meets the safety and reliability requirements of drug recommendations.
Smart Images

Figure CN119049696A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of electronic medical record data processing and auxiliary decision making, and specifically relates to a drug recommendation method, system, device and medium based on deep matrix decomposition. Background Art
[0002] In the past decade, the process of medical digitization has been accelerating. A large number of high-level medical institutions have accumulated rich electronic medical record (EHR) data. Combined with the continuous development of deep learning technology, EHR data has been widely used in medical-related downstream tasks such as mortality prediction, drug recommendation and decision support. In the real clinical diagnosis process, a doctor usually gives a series of drugs for the treatment of the patient's disease; the main task of drug recommendation is to use deep learning technology to mine the patient's health status, past medical history and the corresponding relationship between disease and drug in the electronic medical record data, which can help doctors and medical staff quickly formulate a drug combination that meets the current symptoms according to the patient's specific situation. However, compared with the numerous types of drugs and diseases, the interaction data between drugs and diseases is extremely sparse. In addition, there is the possibility of interaction when drugs are used at the same time. The current recommendation model only focuses on the accuracy of the recommended drug types, but ignores the possible adverse effects, which cannot meet the safety and reliability requirements of drug recommendations.
[0003] Deep learning models have also been widely used in recommendation systems. As one of the most popular technologies in recommendation systems, matrix decomposition is centered on mapping user and item rating matrices into a low-rank space to mine the implicit feature expressions of users and items, and expressing the interaction information between users and items by mapping the inner product of the vectors, thereby predicting the possible relationship between currently unknown users and items. Matrix decomposition relies on the global information in the matrix to fill in the original sparse data. The decomposition process reveals the structure, characteristics, and hidden information of the matrix, has excellent scalability and interpretability, and can alleviate the data sparsity problem of drug recommendations. Based on matrix decomposition and combined with deep learning technology, the matrix recommendation algorithm that integrates deep learning technology is improved and expanded. In existing research, the technology for achieving recommendation purposes based on matrix decomposition has been relatively mature. At present, some studies focus on mining the past medical record information of EHR to obtain more accurate patient feature representation to improve the accuracy of prediction results, and integrate EHR with harmful DDI knowledge graphs, but no one has tried to apply such data to matrix decomposition methods, such as deep matrix factorization (DMF) models. At the same time, existing methods mainly recommend drugs and individual characteristics of patients, but do not pay attention to the overall characteristics of patients and drugs. In addition, the recommendation of drugs has the requirement of safe use of drugs that needs to be considered. On the basis of the effectiveness of the recommended drugs, the probability of adverse drug interactions should be reduced as much as possible, thereby reducing the risk of adverse drug reactions of recommended drugs that endanger patient health.
[0004] The patent application with application number CN202211714897.5 discloses "a deep learning method, system, device and medium for drug recommendation". It uses RNN to establish patient representation, and then integrates the patient's diagnostic information, treatment information, and historical medication information. It uses dual attention to assign weights to different element levels of the patient's historical visits and single visits, and then uses neural networks to make drug recommendations. It can also make drug recommendations based on patient similarity, which effectively improves the accuracy and safety of drug recommendations. However, the application only focuses on the accuracy of the recommended drug types and ignores possible adverse effects. Summary of the invention
[0005] In order to overcome the defects of the above-mentioned prior art, the purpose of the present invention is to propose a drug recommendation method, system, device and medium based on deep matrix decomposition. By constructing a bilateral decomposition method and introducing an attention mechanism and a drug interaction threshold, the model is decomposed from the two perspectives of disease and drug, so as to obtain the potential representations of disease and drug; the introduced attention mechanism distinguishes the attention of the disease to different drug representations to obtain a more accurate preference for the disease; and the drug interaction threshold has certain restrictions on the types of recommended drugs. The drug interaction threshold is reasonably designed to control the drug interaction threshold to avoid the occurrence of adverse drug reactions, thereby effectively improving the safety and practicality of drug recommendation; deep learning is applied to matrix decomposition, and on the basis of the deep decomposition matrix, the potential factors of disease and drug are mined to achieve scientific and safe drug recommendation; deep matrix decomposition uses the nonlinear feature expression ability of deep neural network to replace the simple dot product in the original matrix decomposition to obtain potential factors, and further improves the expression ability and fitting ability of the model by introducing a multi-layer neural network structure, adding activation function, and using batch normalization and other technologies.
[0006] In order to achieve the above object, the technical solution adopted by the present invention is:
[0007] A drug recommendation method based on deep matrix decomposition includes the following steps:
[0008] Step 1: Establish a drug-disease interaction matrix: Use the original data of the disease-drug score matrix to construct an interaction matrix with both explicit scores and non-preference implicit feedback, set thresholds to control drug-drug interactions (DDI), and use the interaction matrix obtained in step 1 as the input of the drug prediction module;
[0009] Step 2: Perform drug prediction tasks: Based on the interaction matrix obtained in step 1, predict the effectiveness of drugs for different diseases, and then recommend drugs based on the obtained therapeutic effectiveness values.
[0010] The specific method of step 1 is:
[0011] 1.1) Clean the electronic medical record data, obtain the relationship between the diseases diagnosed by clinical practice and the corresponding drugs, and obtain the currently known drug-disease correlation values, i.e. the original data R ij , original data R ij The size of can reflect the drug's preference for the disease;
[0012] 1.2) The original data R obtained in step 1.1) is retained, and the correlation value R between the disease and the drug in the original data R is calculated. ij Construct the disease-drug interaction matrix Y to obtain a matrix that retains the original data Rij The disease-drug interaction matrix Y can also contain non-preference implicit feedback; the construction formula of the disease-drug interaction matrix Y is as follows:
[0013]
[0014] Among them, Y ij Represents the association between drug i and disease j in matrix Y. If drug i for disease j has appeared in the same medical record prescription, Y ij =R ij , otherwise Y ij is 0;
[0015] 1.3) For a set of drugs Obtaining drug interaction adjacency matrix from external knowledge The drug interaction adjacency matrix A is traversed to obtain the DDI rate of drug interactions in the actual prescription. The specific formula for calculating the DDI rate of drug interactions in d is as follows:
[0016]
[0017] Among them, A k,j =1 means that drug i and drug j have adverse reactions, DDI d represents the drug-drug interaction DDI rate of drug set d;
[0018] A threshold σ is set that is smaller than the DDI rate of drug-drug interactions in actual prescriptions obtained statistically. The role of the threshold σ is to limit the average DDI rate of drug interactions produced by the recommended drug combinations in the subsequent model training process, so that the average DDI value does not exceed the threshold σ, so as to achieve drug prediction of controllable drug reactions.
[0019] The specific method of step 2 is:
[0020] 2.1) Take the i-th row of the disease-drug interaction matrix Y obtained in step 1 as the disease u i The high-dimensional vector Y i* , high-dimensional vector Y i* The correlation values of all drugs to the i-th disease are expressed, and then the j-th column is taken out as the drug d j The high-dimensional vector Y *j , high-dimensional vector Y *j It is represented by the correlation value of all diseases to the jth drug;
[0021] 2.2) The high-dimensional vector Y of disease and drug obtained in step 2.1) i* , Y *j Input the corresponding multi-layer perceptrons respectively to learn the potential feature representation of diseases and drugs. The input vector l of each layer i, i∈[1,N-1] are converted into low-dimensional outputs of the next layer, where l i =f(W i-1 l i-1 +b i ),i=2,...N-1,W i and b i are the weight values and bias items of each layer respectively;
[0022] 2.3) The ReLU function is used to make the output value of each layer of the multilayer perceptron in step 2.2) non-negative, that is, the positive value of the predicted value is retained, and the negative value is set to 0 to adjust the prediction result. The output result of the last layer of the multilayer perceptron is passed through the ReLU function to obtain the potential feature vector p of the disease and drug. i ,q j ;
[0023] ReLU(x)=max(0,x)
[0024] 2.4) Using the attention mechanism, the latent feature vector q of the drug obtained in step 2.3) is j Input the attention mechanism to calculate the treatment preference of different drug features for various diseases to obtain the attention weight vector Get the attention weight vector Then combine the disease and drug feature vector p i ,q j Calculate the cosine similarity between the two to get the predicted output value of drug j for treating disease i The calculation formula is as follows:
[0025]
[0026] The predicted value The drug output greater than the threshold γ is the final recommended drug result;
[0027] The loss function of the prediction process is expressed as:
[0028] L=αL nce +(1-α)L ddi
[0029]
[0030] A drug recommendation system based on deep matrix decomposition, comprising:
[0031] The data input module uses the original data R of the disease-drug score matrix in step 1. ijConstruct an interaction matrix with both explicit ratings and non-preference implicit feedback, set a threshold to control drug-drug interactions (DDI), and use the interaction matrix obtained in step 1 as the input of the drug prediction module;
[0032] The drug recommendation module, in step 2, predicts the effectiveness of drugs for different diseases based on the interaction matrix obtained in step 1, and recommends drugs based on the obtained therapeutic effectiveness values.
[0033] A drug recommendation device based on deep matrix decomposition, comprising:
[0034] Memory for storing computer programs;
[0035] A processor is used to implement the drug recommendation method based on deep matrix decomposition described in steps 1 to 2 when executing the computer program.
[0036] A computer-readable storage medium is used to store a computer program. When the computer program is executed by a processor, it can make drug recommendations based on the drug recommendation method based on deep matrix decomposition described in steps 1 to 2.
[0037] Compared with the prior art, the present invention has the following beneficial effects:
[0038] 1. Obtain the original association data between diseases and drugs through step 1, and form the interaction matrix between diseases and drugs through the combination of explicit scoring and implicit feedback, which is the initial input of this method. Recommending drugs for diseases is a multi-label binary classification problem. This method processes the original data in accordance with the label characteristics of the problem and can recommend multiple drugs at the same time to achieve more effective, rapid and safe treatment effects.
[0039] 2. Encode the disease and drug into features through step 2, and calculate the low-dimensional feature representation between the disease and drug features through the multi-layer perceptron module. Combine the attention weight to obtain the drug representation that deserves more attention, and use the cosine similarity to obtain the probability prediction value that the drug can treat the disease, so as to determine the recommended drug. The cross entropy loss function for the prediction task in step 2 makes the recommendation more accurate by minimizing the gap between the predicted value and the actual value as much as possible. The loss function for drug interactions limits the incidence of drug interactions, so that the recommended drugs can better calculate the possibility of interaction and ensure the safety of the recommendation task.
[0040] 3. The drug recommendation method based on deep matrix decomposition of the present invention simulates real clinical drug prescriptions and meets the safety and reliability requirements of drug recommendation. It uses deep matrix decomposition with an attention mechanism for the global characteristics of diseases and drugs, introduces nonlinear components to obtain multi-order interaction data, and recommends drugs that have therapeutic effects on diseases. In the disease-drug correlation prediction stage, the existing correlation records between diseases and drugs are used to train the model to obtain low-dimensional feature representations of diseases and drugs. After that, the trained network is used to recommend diseases in the test set to decide whether to recommend the drug prediction module, predict unknown disease-drug relationship pairs, and reconstruct the DMF loss function for drug interactions in drug recommendations, thereby achieving control of the DDI incidence rate, thereby improving the safety of prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0041] Figure 1 This is a schematic diagram of the data input module in step one.
[0042] Figure 2 Schematic diagram of the drug recommendation module in step 2. DETAILED DESCRIPTION
[0043] The present invention is further described in detail below with reference to the accompanying drawings.
[0044] A drug recommendation method based on deep matrix decomposition includes the following steps:
[0045] Step 1: Establish a drug-disease interaction matrix: Use the original data of the disease-drug score matrix to construct an interaction matrix with both explicit scores and non-preference implicit feedback, set thresholds to control drug-drug interactions (DDI), and use the interaction matrix obtained in step 1 as the input of the drug prediction module;
[0046] The specific steps are as follows:
[0047] 1.1) Clean the electronic medical record data, obtain the relationship between the diseases diagnosed by clinical practice and the corresponding drugs, and obtain the currently known drug-disease correlation values, i.e. the original data R ij , R ij The size of can reflect the drug's preference for the disease;
[0048] 1.2) The original data R obtained in step 1.1) is retained, and the correlation value R between the disease and the drug in the original data R is calculated. ij Construct the disease-drug interaction matrix Y to obtain a matrix that retains the original data R ij The disease-drug interaction matrix Y can also contain non-preference implicit feedback; the construction formula of the disease-drug interaction matrix Y is as follows:
[0049]
[0050] Among them, Y ij Represents the association between drug i and disease j in matrix Y. If drug i for disease j has appeared in the same medical record prescription, Y ij =R ij , otherwise Y ij is 0;
[0051] 1.3) For a set of drugs Obtaining drug interaction adjacency matrix from external knowledge The drug interaction adjacency matrix A is traversed to obtain the DDI rate of drug interactions in the actual prescription. The specific formula for calculating the DDI rate of drug interactions in d is as follows:
[0052]
[0053] Among them, A k,j =1 means that drug i and drug j have adverse reactions, DDI d represents the drug-drug interaction DDI rate of drug set d;
[0054] A threshold σ is set that is smaller than the DDI rate of drug-drug interactions in actual prescriptions obtained statistically. The role of the threshold σ is to limit the average DDI rate of drug interactions produced by the recommended drug combinations in the subsequent model training process, so that the average DDI value does not exceed the threshold σ, so as to achieve drug prediction of controllable drug reactions.
[0055] Step 2: Perform drug prediction task: Based on the interaction matrix obtained in step 1, predict the effectiveness of drugs for different diseases, and then recommend drugs based on the obtained therapeutic effectiveness values; the specific steps are as follows:
[0056] 2.1) Take the i-th row of the disease-drug interaction matrix Y obtained in step 1 as the disease u i The high-dimensional vector Y i* , high-dimensional vector Y i* The correlation values of all drugs to the i-th disease are expressed, and then the j-th column is taken out as the drug d j The high-dimensional vector Y *j , high-dimensional vector Y *j It is represented by the correlation value of all diseases to the jth drug;
[0057] 2.2) The high-dimensional vector Y of disease and drug obtained in step 2.1) i* , Y *j Input the corresponding multi-layer perceptrons respectively to learn the potential feature representation of diseases and drugs. The input vector l of each layer i, i∈[1,N-1] are converted into low-dimensional outputs of the next layer, where l i =f(W i-1 l i-1 +b i ),i=2,...N-1,W i and b i are the weight values and bias items of each layer respectively;
[0058] 2.3) The ReLU function is used to make the output value of each layer of the multilayer perceptron in step 2.2) non-negative, that is, the positive value of the predicted value is retained, and the negative value is set to 0 to adjust the prediction result. The output result of the last layer of the multilayer perceptron is passed through the ReLU function to obtain the potential feature vector p of the disease and drug. i ,q j ;
[0059] ReLU(x)=max(0,x)
[0060] 2.4) Using the attention mechanism, the latent feature vector q of the drug obtained in step 2.3) is j Input the attention mechanism to calculate the treatment preference of different drug features for various diseases to obtain the attention weight vector Get the attention weight vector Then combine the disease and drug feature vector p i ,q j Calculate the cosine similarity between the two to get the predicted output value of drug j for treating disease i The calculation formula is as follows:
[0061]
[0062] The predicted value The drug output greater than the threshold γ is the final recommended drug result;
[0063] The loss function of the prediction process is expressed as:
[0064] L=αL nce +(1-α)L ddi
[0065]
[0066] Example
[0067] There is a clinical diagnosis database MIMIC-III, which contains information about patients admitted to the intensive care units of large tertiary care hospitals, including 53,423 different hospital admissions of adult patients to intensive care units between 2001 and 2012. The database MIMIC-III contains 26 data tables, and this embodiment mainly uses the patient medication record table and the patient diagnosis table. The patient diagnosis table contains the patient number and the type of diagnosed disease, and the patient medication record table contains the patient number and the name of the therapeutic drug. The two tables can be merged according to the patient number to obtain a medication data set N. The core of the embodiment of the present invention is to rely on the real clinical diagnosis database MIMIC-III to train a drug recommendation model that can prescribe the right medicine for the right disease. Figure 2 This paper illustrates the process of constructing an interaction matrix with both explicit ratings and non-preference implicit feedback from the medication history N of clinical diagnosis data.
[0068] Step 1: Establish a drug-disease interaction matrix: Use the original data of the disease-drug score matrix to build an interaction matrix with both explicit scores and non-preference implicit feedback, and set thresholds to control drug-drug interactions (DDI). The specific steps are as follows:
[0069] 1.1) Clean the above medication history dataset N, filter out empty data in the dataset, remove the patient number information column, merge the same disease-drug rows, count the number of occurrences of each disease-drug row during the merging process, add a new drug-disease relationship column, and fill the newly added drug-disease relationship column with the number of occurrences. The more the number of occurrences of the disease-drug row, the higher the possibility that the drug is used to treat this type of disease. In this way, data on whether the drug has a therapeutic relationship with the disease through clinical diagnosis is obtained. According to the drug-disease relationship column, the currently known drug-disease correlation value is analyzed, that is, the original data R ij , R ij The size reflects the drug's preference for the disease;
[0070] 1.2) The original data R obtained in step 1.1) ij Retain the original data R containing the currently known correlation values between diseases and drugs ij Construct a global disease-drug interaction matrix Y that includes both explicit preferences and non-preference implicit feedback; the formula for constructing the disease-drug interaction matrix Y is as follows:
[0071]
[0072] For ease of understanding, Figure 2 The construction process of the disease-drug interaction matrix Y shown in the figure is only an abstract summary. The actual process requires the original data R ijTraverse to fill the matrix Y.
[0073] 1.3) For a set of drugs Obtaining drug interaction adjacency matrix from external knowledge A i,j =1 indicates that drug i and drug j have adverse reactions. The drug interaction adjacency matrix A is traversed to obtain the DDI rate of drug interactions in the real diagnostic data. The specific formula for calculating the DDI rate of drug interactions in d is as follows:
[0074]
[0075] In order to ensure that the recommended drug combination in the subsequent model training process is safe and reliable, it is necessary to limit the average DDI rate of drug interactions, and set a threshold σ that is slightly smaller than the statistically obtained DDI rate of drug interactions in real diagnosis, so that the average DDI value does not exceed the threshold σ, so as to achieve drug prediction of controllable drug reactions. Assuming that the probability of drug interaction in real diagnosis in this embodiment is 0.08, the threshold σ can be set to a value less than 0.08. The smaller the setting, the higher the theoretical safety of the recommended drug, and the more the recommended performance is restricted.
[0076] In this embodiment, after obtaining the input data disease-drug interaction matrix Y and threshold σ of the model, the input Figure 1 The drug recommendation module shown starts training the model.
[0077] Step 2: Perform drug prediction task: Use the interaction matrix Y obtained in step 1 as the input of the drug prediction module to predict the effectiveness of drugs for different diseases, so as to recommend drugs based on the obtained therapeutic effectiveness values; the specific steps are as follows:
[0078] 2.1) Take the i-th row of the disease-drug interaction matrix Y obtained in step 1 as the disease u i The high-dimensional vector Y i* , Y i* The correlation values of all drugs to the i-th disease are expressed, and then the j-th column is taken out as the drug d j The high-dimensional vector Y *j , Y *j It is represented by the correlation value of all diseases to the jth drug;
[0079] 2.2) The high-dimensional vector Y of disease and drug obtained in step 2.1) i* , Y *j Input the corresponding multi-layer perceptrons respectively to learn the potential feature representation of diseases and drugs. The input vector l of each layer in the perceptron i , i∈[1,N-1] are converted into low-dimensional outputs of the next layer, where li =f(W i-1 l i-1 +b i ),i=2,...N-1,W i and b i are the weight values and bias items of each layer respectively; in this embodiment, the best effect is achieved when the model has two layers;
[0080] 2.3) The ReLU function is used to make the output value of each layer of the multilayer perceptron in step 2.2) non-negative, that is, the positive value of the predicted value is retained, and the negative value is set to 0 to adjust the prediction result. The output result of the last layer of the multilayer perceptron is passed through the ReLU function to obtain the potential feature vector p of the disease and drug. i ,q j ;
[0081] ReLU(x)=max(0,x)
[0082] 2.4) Using the attention mechanism, the latent feature vector q of the drug obtained in step 2.3) is j Input the attention mechanism to calculate the treatment preference of different drug features for various diseases to obtain the attention weight vector Get the attention weight vector Then combine the disease and drug feature vector p i ,q j Calculate the cosine similarity between the two to get the predicted output value of drug j for treating disease i The calculation formula is as follows:
[0083]
[0084] The predicted value The drug output greater than the threshold γ is used to obtain the final recommended drug result. In this embodiment, after the model training is completed, any disease input in the medication data set N can obtain the corresponding recommended drug, for example, ibuprofen is recommended for fever.
[0085] The loss function of the prediction process is expressed as:
[0086] L=αL nce +(1-α)L ddi
[0087]
[0088] In this embodiment, the model training is set to 200 iterations until the model's recommendation accuracy reaches a relatively stable state, and the model training is completed. The number of iterations of model training can be changed according to the initial input data situation. This embodiment is only for the medication history data set N.
[0089] The drug recommendation method based on deep matrix decomposition of the present invention simulates real clinical drug prescriptions and meets the safety and reliability requirements of drug recommendation. It uses deep matrix decomposition with an attention mechanism for the global characteristics of diseases and drugs, introduces nonlinear components to obtain multi-order interaction data, and recommends drugs that have therapeutic effects on diseases. In the disease-drug correlation prediction stage, the existing correlation records between diseases and drugs are used to train the model to obtain low-dimensional feature representations of diseases and drugs. After that, the trained network is used to recommend diseases in the test set to decide whether to recommend the drug prediction module, predict unknown disease-drug relationship pairs, and reconstruct the DMF loss function for drug interactions in drug recommendations, thereby achieving control of the DDI incidence rate, thereby improving the safety of prediction.
Claims
1. A drug recommendation method based on deep matrix decomposition, comprising the following steps: Step 1: Establish a drug-disease interaction matrix: Use the original data of the disease-drug score matrix to construct an interaction matrix with both explicit scores and non-preference implicit feedback, set thresholds to control drug-drug interactions (DDI), and use the interaction matrix obtained in step 1 as the input of the drug prediction module; Step 2: Perform drug prediction tasks: Based on the interaction matrix obtained in step 1, predict the effectiveness of drugs for different diseases, and then recommend drugs based on the obtained therapeutic effectiveness values.
2. The drug recommendation method based on deep matrix decomposition according to claim 1, characterized in that: The specific method of step 1 is: 1.1) Clean the electronic medical record data, obtain the relationship between the diseases diagnosed by clinical practice and the corresponding drugs, and obtain the currently known drug-disease correlation values, i.e. the original data R ij , R ij The size of can reflect the drug's preference for the disease; 1.2) The original data R obtained in step 1.1) is retained, and the correlation value R between the disease and the drug in the original data R is calculated. ij Construct the disease-drug interaction matrix Y to obtain a matrix that retains the original data R ij The disease-drug interaction matrix Y can also contain non-preference implicit feedback; the construction formula of the disease-drug interaction matrix Y is as follows: Among them, Y ij represents the association between drug i and disease j in matrix Y; if drug i for disease j has appeared in the same medical record prescription, Y ij =R ij , otherwise Y ij is 0; 1.3) For a set of drugs Obtaining drug interaction adjacency matrix from external knowledge The drug interaction adjacency matrix A is traversed to obtain the DDI rate of drug interactions in the actual prescription. The specific formula for calculating the DDI rate of drug interactions in d is as follows: Among them, A k,j =1 means that drug i and drug j have adverse reactions, DDI d represents the DDI rate of drug interactions among drug set d; A threshold σ is set that is smaller than the DDI rate of drug-drug interactions in actual prescriptions obtained statistically. The role of the threshold σ is to limit the average DDI rate of drug interactions produced by the recommended drug combinations in the subsequent model training process, so that the average DDI value does not exceed the threshold σ, so as to achieve drug prediction of controllable drug reactions.
3. The drug recommendation method based on deep matrix decomposition according to claim 1, characterized in that: The specific method of step 2 is: 2.1) Take the i-th row of the disease-drug interaction matrix Y obtained in step 1 as the disease u i The high-dimensional vector Y i* , high-dimensional vector Y i* The correlation values of all drugs to the i-th disease are expressed, and then the j-th column is taken out as the drug d j The high-dimensional vector Y *j , high-dimensional vector Y *j It is represented by the correlation value of all diseases to the jth drug; 2.2) The high-dimensional vector Y of disease and drug obtained in step 2.1) i* , Y *j Input the corresponding multi-layer perceptrons respectively to learn the potential feature representation of diseases and drugs. The input vector l of each layer i , i∈[1,N-1] are converted into low-dimensional outputs of the next layer, where l i =f(W i-1 l i-1 +b i ),i=2,...N-1,W i and b i are the weight values and bias items of each layer respectively; 2.3) The ReLU function is used to make the output value of each layer of the multilayer perceptron in step 2.2) non-negative, that is, the positive value of the predicted value is retained, and the negative value is set to 0 to adjust the prediction result. The output result of the last layer of the multilayer perceptron is passed through the ReLU function to obtain the potential feature vector p of the disease and drug. i ,q j ; ReLU(x)=max(0,x) 2.4) Using the attention mechanism, the latent feature vector q of the drug obtained in step 2.3) is j Input the attention mechanism to calculate the treatment preference of different drug features for various diseases to obtain the attention weight vector Get the attention weight vector Then combine the disease and drug feature vector p i ,q j Calculate the cosine similarity between the two to get the predicted output value of drug j for treating disease i The calculation formula is as follows: The predicted value The drug output greater than the threshold γ is the final recommended drug result; The loss function of the prediction process is expressed as: L=αL nce +(1-α)L ddi 4. A drug recommendation system based on the method according to any one of claims 1 to 3, characterized in that: include: The data input module uses the original data R of the disease-drug score matrix in step 1. ij Construct an interaction matrix with both explicit ratings and non-preference implicit feedback, set a threshold to control drug-drug interactions (DDI), and use the interaction matrix obtained in step 1 as the input of the drug prediction module; The drug recommendation module, in step 2, predicts the effectiveness of drugs for different diseases based on the interaction matrix obtained in step 1, and recommends drugs based on the obtained therapeutic effectiveness values.
5. A drug recommendation device based on the method according to any one of claims 1 to 3, characterized in that: A drug recommendation device based on deep matrix decomposition, comprising: Memory for storing computer programs; A processor is used to implement the drug recommendation method based on deep matrix decomposition described in steps 1 to 2 when executing the computer program.
6. A computer-readable storage medium for storing a computer program, characterized in that: When the computer program is executed by a processor, it is capable of performing drug recommendations based on the drug recommendation method based on deep matrix decomposition as described in claims 1 to 3.
Citation Information
Cited By
Drug recommendation system and method based on causal inference
CN119811576A
