A deep learning method, system, device, and medium for drug recommendation

Abstract

A deep learning method, system, device, and medium for drug recommendation are disclosed. The method includes: establishing drug representations; establishing patient representations; and generating drug prediction results. These three steps first utilize an RNN to establish patient representations, then integrate patient diagnostic information, treatment information, and historical medication information, employing dual attention to assign weights at different element levels in patient history and single visits, and finally using a neural network for drug recommendation. Furthermore, patient similarity can be incorporated for drug recommendation, effectively improving the accuracy and safety of drug recommendations. The system, device, and medium, through the storage and utilization of relevant functional modules, realize drug recommendation using deep learning methods. As a medical auxiliary tool, it can significantly improve the discrimination quality and work efficiency of pathologists.

Description

Technical Field

[0001] This invention belongs to the field of drug-assisted decision-making technology, specifically relating to a deep learning method, system, device, and medium for drug recommendation. Background Technology

[0002] Drug recommendation is an important application of artificial intelligence in the medical field. The goal of drug recommendation is to suggest appropriate drug combinations based on a patient's current treatment methods and diagnostic results, combined with their medical history, to assist doctors in the treatment process. However, drug interactions exist, and recommending effective and safe drug combinations is particularly challenging when a patient's condition is complex.

[0003] Recurrent Neural Networks (RNNs) are a series of deep learning models that learn from sequential and time-series data. Attention mechanisms assign different weights to different elements in a sequence based on the inherent relevance of the sequence. Larger weights indicate a greater relevance between the corresponding component and the final representation. Existing methods recommend medications based on a patient's current diagnosis, using content attention mechanisms and a multi-instance, multi-label learning framework to predict drug combinations end-to-end based on the patient's diagnosis. However, this method neglects the patient's historical medical history and changes in their condition over time. Currently, some studies are beginning to utilize RNNs to establish a dual-attention mechanism, assigning different weights to the patient's diagnosis and treatment methods in different visits and single visits, providing a comprehensive representation of the patient. However, these methods do not consider the impact of past medication use on the patient, nor do they utilize patient similarity to optimize medication recommendations.

[0004] The drug recommendation method disclosed in patent application CN111753543A uses reinforcement learning to recommend drugs according to the actor-critic paradigm by obtaining relevant information about the target object and using a subsequent model one for drug prediction, and then using a subsequent model two to provide feedback. This method requires training two models, model one and model two, and requires setting two types of rewards, making the model architecture relatively complex. Summary of the Invention

[0005] To overcome the shortcomings of the existing technologies, the present invention aims to propose a deep learning method, system, device, and medium for drug recommendation. By establishing a comprehensive patient representation through a dual attention mechanism, a more comprehensive patient representation is obtained. At the same time, the similarity between patients is used to optimize drug prediction results, and the drug and drug response thresholds are controlled, effectively improving the accuracy of drug recommendation. By constructing a neural network for drug prediction, the training process can be significantly accelerated by utilizing existing patient information data and a simple model architecture, while ensuring the effectiveness of drug recommendation.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A deep learning method for drug recommendation includes the following steps:

[0008] Step 1: Establishing Drug Representation: Establish drug representation using graph convolutional neural network, and set thresholds to control drug and drug response;

[0009] 1) Use graph convolutional neural networks to process drug representations, for drug embedding vectors... Drug adjacency matrix corresponding to the drug , This represents the total number of drugs; the graph convolutional neural network will process the drugs according to the following formula to obtain a new drug embedding vector:

[0010]

[0011] in , It is the degree matrix of A, i.e. ;

[0012] 2) Based on the convolutional neural network method in step 1), initialize the drug representation. Clinical co-occurrence matrix Drug Interaction Matrix Performing double GCN processing separately is represented as follows:

[0013]

[0014]

[0015] Where ReLU is the activation function; and These are all learnable parameters. Indicates drug With drugs It appeared in the same prescription before. Indicates drug With drugs Adverse reactions exist; the results are subtracted, and the difference is used as a representation of the effects of GCN treatment on different drugs, i.e.: ,in , Each row of elements represents the embedding vector of the corresponding drug after processing using the clinical co-occurrence matrix and the drug interaction matrix; for The modeling is represented as follows:

[0016]

[0017] in ; M Indicates medicine;

[0018] For a group of drugs The DDI value of a drug is expressed as:

[0019]

[0020] During model training, by setting a threshold Controllable drug response prediction can be achieved by limiting the average DDI value of each patient's visits;

[0021] Step 2: Establishing the patient representation: Combining the patient's historical state and current state, a comprehensive patient representation is established using the patient's diagnostic information, treatment methods, and medication information. Based on the dual attention mechanism, the patient's historical state representation and current state representation are spliced ​​together to obtain the final patient state representation, thereby obtaining a more comprehensive patient representation.

[0022] Step 3, Drug Prediction Task: Based on the patient representation learned in Step 2, the data is directly input into a single-layer network for drug prediction, or the similarity between the patient and other patients is compared to obtain an optimized patient representation using the similarity between patients, thereby improving the prediction results.

[0023] The specific method for step two is as follows:

[0024] 1) Represent a patient's hospital visit experience as a sequence. , where t represents the number of times the patient visits the hospital; a single visit by the patient. Diagnosis by 0-1 vector Diagnostic and treatment methods ,drug Composition; of which , , These represent the total number of diagnoses, the total number of treatments, and the total number of medications, respectively. By concatenating three sets of 0-1 vectors, we can obtain the 0-1 vector of a patient's visit. Initialize an embedding vector table The patient's diagnosis, treatment methods, and medications can be projected into the corresponding embedding space using the inner product, as shown below:

[0025]

[0026] in, The composite embedding vector used to represent the patient is used, and the embedding table E contains the parameters to be learned; based on the modeling results of step 2) of step one. The embedding vectors of the medications obtained during a patient's visit are summed, i.e., a patient's visit i is represented as:

[0027]

[0028] The patient's last visit was indicated as , where 0…0 represents The zero vector; therefore, the patient's last visit Represented as:

[0029]

[0030] in ;

[0031] 2) Based on the patient's i-th visit information obtained in step 1) of step two, The historical visit sequence of patients with a total of t visits (t>1) is denoted as follows: ,in Input the patient's historical visit sequence The data obtained included patients' statements from each visit, along with their historical information.

[0032]

[0033] The patient's history of visits is represented as follows: ,in ;

[0034] 3) Using the attention mechanism, input the patient history visit sequence obtained in step 2) of step two into two different attention mechanisms. and Each is assigned a different weight. The calculation process is as follows:

[0035]

[0036]

[0037] in It is a learnable parameter vector;

[0038]

[0039]

[0040] in , It is a learnable parameter vector; through the two attention mechanisms mentioned above, a new representation of each patient's historical visit record is obtained:

[0041]

[0042] in, , This represents element-wise multiplication between two vectors;

[0043] The total patient visit history is obtained by summing the patient's t-1 historical visit records as follows:

[0044] ,

[0045] Representing the patient's total visit history and the patient's current health status. By splicing the images together, we obtain the final representation of the patient's current state:

[0046]

[0047] in .

[0048] The specific method for step three is as follows:

[0049] The final representation of the patient's current state is obtained from step 3 of step 2. Then, it is directly input into a single-layer neural network. After obtaining the corresponding output, it is then passed through a sigmoid function to obtain the final drug prediction output for the patient's t-th visit, as shown below:

[0050]

[0051] Optimize drug recommendations based on patient similarity and calculate the final representation of the patient's current state. With patient sequence storage block The cosine similarity between the sequence representations of other patients is calculated as follows:

[0052]

[0053] in Representing the sequence of the patient's medical history and Cosine similarity between them; using cosine similarity of patient medical history sequences The calculation process for aggregating similar patient sequences as attention weights is as follows:

[0054]

[0055] in This represents the current patient representation, integrating sequences of similar patient medical histories.

[0056] Input it into a single-layer neural network After obtaining the corresponding output, it is then passed through a sigmoid function to obtain the final drug prediction output for the patient's t-th visit, resulting in an improved drug recommendation result. It is expressed as follows:

[0057] .

[0058] A deep learning system for drug recommendation, implementing the deep learning methods described in steps one to three, includes:

[0059] The patient representation module uses the patient's current diagnostic information, treatment methods, and past diagnostic information, treatment methods, and medication history to obtain the final patient representation by integrating historical information and visit sequence;

[0060] The drug representation module utilizes drug co-occurrence information and drug response information from prescription history to obtain a graph convolutional neural network-enhanced drug representation;

[0061] The drug recommendation module obtains the final drug recommendation result by choosing between comparing patient history similarity or directly using a single-layer neural network.

[0062] A deep learning device for drug recommendation includes:

[0063] Memory, used to store data related to computer programs and drug recommendations;

[0064] A processor is used to implement the deep learning method for drug recommendation described in steps one through three when executing the computer program.

[0065] A computer-readable storage medium storing a computer program that, when executed by a processor, enables drug recommendation using the deep learning methods described in steps one to three.

[0066] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0067] The method of this invention establishes a patient's drug representation in step one. Leveraging the powerful graph representation capabilities of GCNs, a graph convolutional neural network is used to process the drug representation. Considering co-occurrence and interactions between drugs can improve the safety and reliability of drug recommendations. In step two, an RNN is used to establish a representation integrating the patient's medication history, visit information, and treatment methods. A dual attention encoder is then used to obtain a comprehensive representation of the patient across the visit time span, attempting to identify key diagnostic and treatment approaches. Step three utilizes a single-layer neural network to obtain the final drug prediction output. Furthermore, the drug recommendation process can be optimized by comparing patient similarity.

[0068] These three steps primarily consider drug characteristics, patient history of visits, and current diagnostic status, establishing a comprehensive patient profile and effectively improving the accuracy and safety of drug recommendations.

[0069] This invention uses a neural network for drug prediction. By utilizing existing patient information data and a simple model architecture, the training process can be significantly accelerated while ensuring the effectiveness of drug recommendations. Attached Figure Description

[0070] Figure 1 This is a framework diagram of the drug recommendation method based on two-layer attention proposed in this invention.

[0071] Figure 2 This is a schematic diagram of the basic attention mechanism of the present invention.

[0072] Figure 3 This is a schematic diagram of the dual attention mechanism of the present invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings.

[0074] See Figure 1 A deep learning method for drug recommendation includes the following steps:

[0075] Step 1: Establishing Drug Representation: Drug representation is established using a graph convolutional neural network. A threshold is set to control drug-drug interactions (DDI), allowing the model to be trained while controlling the upper limit of DDI, thus achieving drug prediction of controllable drug responses.

[0076] 1) Process drug representations using existing graph convolutional neural networks. For a given drug embedding vector... Drug adjacency matrix corresponding to the drug The graph convolutional neural network will process the drug according to the following formula to obtain a new drug embedding vector:

[0077]

[0078] in , It is the degree matrix of A, i.e. ;

[0079] 2) Based on the convolutional neural network method in step 1), initialize the drug representation. Clinical co-occurrence matrix Drug Interaction Matrix The structure of the double GCN can be represented as follows:

[0080]

[0081]

[0082] in, and These are all learnable parameters. Indicates drug With drugs It appeared in the same prescription before. Indicates drug With drugs Adverse reactions exist. The results are subtracted, and the difference is used to represent the effects of different drugs after GCN treatment. That is: ,in Each row of elements represents the embedding vector of the corresponding drug after processing using the clinical co-occurrence matrix and the drug interaction matrix; it is modeled as follows:

[0083]

[0084] in, .

[0085] Step Two: Establishing the Patient Representation: Combining historical and current situations, a comprehensive patient representation is established. This representation is obtained by concatenating the patient's historical state representation and current state representation based on a dual attention mechanism, thus obtaining a more comprehensive patient representation.

[0086] 1) Represent a patient's hospital visit experience as a sequence. , where t represents the number of times the patient visits the hospital. A single visit by the patient. Diagnosis by multidimensional vector Diagnostic and treatment methods ,drug Composition. Among them... , , These represent the total number of diagnoses, the total number of treatments, and the total number of medications, respectively. Concatenating these three-dimensional elements yields a multi-dimensional 0-1 vector representing a patient's specific visit. Initialize an embedding vector table. The patient's diagnosis, treatment methods, and medications can be projected into the corresponding embedding space using the inner product, which can be represented as:

[0087]

[0088] in, The composite embedding vector is used to represent the patient, and the embedding table E contains the parameters to be learned. This is based on the modeling results from step 2) of step one. The embedding vectors of the medications obtained during a patient's visit are summed, meaning that a patient's visit i can be represented as:

[0089]

[0090] The patient's last visit can be expressed as , where 0…0 represents The zero vector. Therefore, the patient's last visit It can be represented as:

[0091]

[0092] in ;

[0093] 2) Based on the patient's i-th visit information obtained in step 1) of step two, Then, the historical visit sequence of patients with a total of t visits (t>1) can be denoted as: ,in Input the patient's historical visit sequence From this, we can obtain the patient's visit history, including historical information:

[0094]

[0095] The patient's history of visits is represented as follows: ,in ;

[0096] 3) Using the attention mechanism, input the patient history visit sequence obtained in step 2) of step two into two different attention mechanisms. and Each is assigned a different weight. The calculation process is as follows:

[0097]

[0098]

[0099] in It is a learnable parameter vector;

[0100]

[0101]

[0102] in , It is a learnable parameter vector; through the two attention mechanisms mentioned above, a new representation of each patient's historical visit record is obtained:

[0103]

[0104] in , This represents element-wise multiplication between two vectors;

[0105] The total patient visit history is obtained by summing the patient's t-1 historical visit records as follows:

[0106] ,

[0107] The patient's total visit history is represented by the patient's current health status. By splicing the images together, we obtain the final representation of the patient's current state:

[0108]

[0109] in, .

[0110] Step 3, Drug Prediction Task: After learning the patient representation in Step 2, the data can be directly input into a single-layer network for drug prediction. Alternatively, the similarity between patients can be compared to that of other patients to obtain an optimized patient representation, thus improving the prediction results.

[0111] The final representation of the patient's current state is obtained from step 3 of step 2. Then, it can be directly input into a single-layer neural network. After obtaining the corresponding output, it is then passed through a sigmoid function to obtain the final drug prediction output for the patient's t-th visit, as shown below:

[0112]

[0113] If patient similarity is used to optimize drug recommendations, the final representation of the patient's current state can be calculated. With patient sequence storage block The cosine similarity between the sequence representations of other patients is calculated as follows:

[0114]

[0115] in Representing the sequence of the patient's medical history and Cosine similarity between them. Cosine similarity using patient medical history sequences. The calculation process for aggregating similar patient sequences as attention weights is as follows:

[0116]

[0117] in This represents the current patient representation, integrating sequences of similar patient medical histories.

[0118] Input it into a single-layer neural network After obtaining the corresponding output, it is then passed through a sigmoid function to obtain the final drug prediction output for the patient's t-th visit, resulting in an improved drug recommendation result. It is expressed as follows:

[0119]

Claims

1. A deep learning method applied to drug recommendation, characterized in that: Specifically comprising the following steps: Step one, the establishment of drug representation: the establishment of drug representation through graph convolutional neural network, setting threshold to control drug and drug reaction; 1) processing drug representations using a graph convolutional neural network, for drug embedding vectors and a drug adjacency matrix corresponding to the drugs , representing the total number of drugs; the graph convolutional neural network will process the drugs to obtain new drug embedding vectors according to the following equation: wherein , is the degree matrix of A, i.e. ; 2) Initialize drug representations according to the graph convolutional neural network method in step one, 1) Through the clinical co-occurrence matrix With the drug interaction matrix Respectively, the double GCN processing is represented as: where ReLU is the activation function; are all learnable parameters, denotes the drug denotes the drug denotes the drug have appeared in the same prescription, denotes the drug denotes the drug There are adverse reactions; the difference between the results obtained is used as the representation of different drugs after GCN processing, that is: where , Each row element in represents the embedding vector of the corresponding drug after processing the clinical co-occurrence matrix and the drug interaction matrix; modeling is represented as: wherein ; M denotes a drug; For a set of drugs , the DDI value for a drug is expressed as: During the model training process, by setting a threshold Limiting the average DDI value of each patient's visit can achieve drug prediction of controllable drug reactions; Step two, the establishment of patient representation: combining the patient's historical state and current state, using the patient's diagnosis information, diagnosis and treatment methods and drug information to establish the comprehensive representation of the patient, based on the double attention mechanism to splice the patient's historical state representation and the patient's current state representation as the final patient state representation, so as to obtain a more comprehensive patient representation; Step three, drug prediction task: after learning the patient representation according to step two, directly inputting single layer network for drug prediction, or comparing the similarity between the patient and other patients, using the similarity between patients to obtain the optimized patient representation, and improving the prediction result. 2.The deep learning method applied to drug recommendation according to claim 1, characterized in that: The specific method of step two is: 1) The visit experience of a patient to a hospital is represented as a sequence , where t is the number of visits of the patient to the hospital; the visit experience of a patient at a time t is diagnosed by a 0-1 vector , the treatment means , and the drugs ; where , , respectively represent the total number of diagnoses, the total number of treatment means, and the total number of drugs; by splicing the three 0-1 vectors, a 0-1 vector of a patient at a visit t is obtained; an embedding vector table is initialized , and by inner product, the diagnosis, treatment means, and drugs of a patient can be projected into the corresponding embedding space, represented as: wherein, is used to represent the patient's composite embedding vector, the embedding table E is a parameter to be learned; the result of modeling according to step one, 2) step The embedding vectors of the drugs obtained during the patient's visit i are added, i.e. the patient's visit i is represented as: The last visit of the patient is denoted as where 0…0 denotes the zero vector of ; thus the last visit of the patient is denoted as : wherein ; 2) the patient's i-th visit representation obtained from step two, 1) above The patient's history visit sequence with t visits (t > 1) is denoted as where The patient's history visit sequence is input into to obtain the patient's visit representations with history information The patient's historical visit sequence is represented as , wherein ; 3) Using attention mechanism, input the patient history visit sequence obtained in step two 2) into two different attention mechanisms respectively With , each is given a different weight The calculation process is as follows: wherein is a learnable parameter vector; wherein , is a learnable parameter vector; through the above two attention mechanisms, a new representation of each historical visit record of the patient is obtained: wherein , denotes an element-wise multiplication between two vectors; Add the patient's t-1 historical visit record to obtain the patient's total visit history representation: , representations of the patient's current health state performing the stitching to obtain a final representation of the patient's current state: wherein . 3.The deep learning method applied to drug recommendation of claim 1, wherein: The specific method of step three is: According to step two, the third step, the final representation of the current state of the patient is obtained Then, it is directly input into a single-layer neural network After obtaining the corresponding output, it is passed through a sigmoid function to obtain the final drug prediction output for the tth visit of the patient, which is represented as follows: Selecting patient similarity to optimize drug recommendation results, computing the final representation of this patient's current state Cosine similarity between patient sequence representations and other patient sequence representations in the patient sequence store block The cosine similarity between patient sequence representations and other patient sequence representations in the patient sequence store block, the computation flow is as follows: wherein representing a patient history sequence and cosine similarity between them; using patient history sequence cosine similarity as attention weights to aggregate similar patient sequences, the computational flow of which is as follows: wherein a current patient representation representing a sequence of integrated similar patient histories; Inputting it into a single-layer neural network , and then passing it through a sigmoid function to obtain the final drug prediction output for the tth visit of the patient, to obtain an improved drug recommendation result is expressed as follows: 。 4. A deep learning system applied to drug recommendation, characterized by: The deep learning method of any one of claims 1 to 3 comprises: A patient representation module that uses the patient's current diagnosis information, diagnosis and treatment methods, and the patient's past diagnosis information, diagnosis and treatment methods, and medication history to obtain the final representation of the patient fused with historical information and visit sequence; A drug representation module that uses the drug co-occurrence information and drug reaction information in the prescription history to obtain the graph convolutional neural network enhanced drug representation; A drug recommendation module that obtains the final drug recommendation result by selecting the patient historical similarity comparison or directly through a single layer neural network. 5.A deep learning device applied to drug recommendation, characterized in that: It comprises: A memory for storing computer programs and related data for drug recommendation; A processor for executing the computer program to realize the deep learning method of any one of claims 1 to 3 applied to drug recommendation.

6. A computer-readable storage medium, characterized in that: The computer readable storage medium stores a computer program, which can be used for drug recommendation by the deep learning method of any one of claims 1 to 3 when executed by the processor.

Images