Drug recommendation method and system that refer to health conditions similar to health condition of patient without distinguishing patient

Abstract

A drug recommendation method and system that refer to health conditions similar to the health condition of a patient without distinguishing the patient are disclosed. The drug recommendation method performed by the drug recommendation system, according to one embodiment, may comprise the steps of: using diagnosis history information and surgery history information obtained during the current visit of a patient and the past visit of the patient, thereby acquiring a patient representation indicating the current health condition of the patient; and using first drug history information reflecting the degree to which drugs are prescribed together, and second drug history information obtained on the basis of past health conditions of a patient, who is similar to the patient with respect to the current health condition, and past health conditions of other patients, thereby recommending drugs to the patient on the basis of the acquired patient representation.

Description

A method and system for recommending medicines that refer to health conditions similar to the patient's health condition without distinction between patients.

[0001] The description below is about drug recommendation techniques.

[0002] Drug recommendation research has been actively researched because it can save physicians the time and effort required to prescribe accurate medications and can be useful even in situations where physicians are understaffed. Early drug recommendation methods recommended medications based solely on the patient's diagnosis or surgical history at the current visit. However, because these methods did not consider the patient's past health history, they could not distinguish whether the diagnosis or surgical procedure prescribed at the current visit was acute or chronic. As a result, they recommended the same medications to different patients with the same diagnosis or surgical history at the current visit, limiting their ability to accurately recommend medications to patients.

[0003] Recently, medication recommendation methods have begun to consider not only the diagnosis / surgical history prescribed at the current visit but also past health records to alleviate the limitations of conventional methods. While these methods demonstrate greater accuracy than conventional methods by leveraging patients' past health records, they struggle to utilize only medications prescribed at past visits that correspond to a similar health status to the patient's current condition, regardless of whether the medications were prescribed by the patient or by others. Furthermore, EHR (Electronic Health Records) graphs fail to accurately represent the degree to which each medication should be prescribed alongside the others, even if the two medications were previously prescribed together.

[0004] After flexibly selecting health states similar to the patient's current health state without distinction between patients, we utilize the prescription history of medications prescribed in the selected health state, and for two medications that have been prescribed together, we represent them as edges with directions pointing toward each other from the perspective of each medication, and then consider the degree to which each medication should be prescribed together with the remaining medications as the weight of each edge to recommend medications that are effective for the patient.

[0005] A method for recommending a drug performed by a drug recommendation system may include the steps of: obtaining a patient expression representing the current health state of the patient by using the diagnosis history information and surgical history information prescribed at the patient's current visit and the patient's past visits; and recommending a drug to the patient by using first drug history information reflecting the degree of prescription among drugs based on the obtained patient expression, and second drug history information prescribed at a past health state of a patient similar to the current health state of the patient and at a past health state of another patient.

[0006] The above recommended step may include a step of obtaining a drug representation by considering relationship information between drugs prescribed together and relationship information that is less likely to cause side effects when taken together, through EHR graph+ and DDI (Drug Drug Interaction) graph.

[0007] The above recommended steps may include using EHR Graph+ to represent all drugs as directed edges for each drug pair, and assigning a weight to the edge based on the degree to which the drugs are prescribed together at each drug level.

[0008] The above recommended step may include a step of obtaining a first medication expression required for the patient's current visit based on the obtained medication expression and the obtained patient expression.

[0009] The above-described recommending step may include a step of obtaining a second medication expression required for the patient's current visit by reflecting the degree to which each medication is related to the patient expression of the current visit and the patient expression of the past visits in which the medication was prescribed, based on the patient expression obtained through Health Status-Aware Attention and medications prescribed in past visits showing other patient expressions similar to the obtained patient expression.

[0010] The above recommended step may include a step of deriving the medications required for the patient's current visit by fusing the first medication expression and the second medication expression.

[0011] The above Health Status-Aware Attention trains a drug recommendation model using a learning set, then performs hyperparameter tuning of the trained drug recommendation model using a verification set, obtains patient expressions for all visits of all patients from the drug recommendation model confirmed through the hyperparameter tuning, and selects a specific number of visits higher than a preset rank among similar past visits based on the obtained patient expressions based on cosine similarity.

[0012] The drug recommendation system may include a patient expression acquisition unit that acquires a patient expression representing the current health status of the patient by using the diagnosis history information and surgical history information prescribed at the patient's current visit and the patient's past visits; and a drug recommendation unit that recommends a drug to the patient by using first drug history information reflecting the degree of prescription among drugs based on the acquired patient expression and second drug history information prescribed at a past health status of a patient similar to the current health status of the patient and at a past health status of another patient.

[0013] Unlike conventional technologies, when utilizing medication records prescribed in past visits, the use is not limited to medication records prescribed in the patient's past visits, but can only utilize medication records prescribed in the past health conditions of the patient or other patients similar to the patient's current health condition.

[0014] All drugs are nodes, and if any two drugs have been prescribed together in at least one of all patient visits, simply connecting the two drug nodes with an edge, that is, unlike conventional techniques that utilize EHR graphs without weights or directions, it is possible to obtain a drug representation that accurately reflects the degree of co-prescription between two drugs that have been prescribed together.

[0015] This allows us to recommend safe medicines within the range of acceptable prescriptions while improving accuracy.

[0016] FIG. 1 is a block diagram illustrating the configuration of a drug recommendation system in one embodiment.

[0017] FIG. 2 is a flowchart illustrating a method for recommending a drug in one embodiment.

[0018] FIG. 3 is a diagram for explaining a drug recommendation operation in one embodiment.

[0019] Hereinafter, the detailed description will be given with reference to the attached drawings.

[0020] FIG. 1 is a block diagram for explaining the configuration of a drug recommendation system in one embodiment, and FIG. 2 is a flowchart for explaining a drug recommendation method in one embodiment.

[0021] The processor of the drug recommendation system (100) may include a patient expression acquisition unit (110) and a drug recommendation unit (120). These components of the processor may be representations of different functions performed by the processor according to control commands provided by program codes stored in the drug recommendation system. The processor and the components of the processor may control the drug recommendation system that performs steps (210 to 220) included in the drug recommendation method of FIG. 2. At this time, the processor and the components of the processor may be implemented to execute instructions according to the code of the operating system included in the memory and the code of at least one program.

[0022] The processor can load program code stored in a program file for a drug recommendation method into memory. For example, when a program is executed in a drug recommendation system, the processor can control the drug recommendation system to load the program code from the program file into memory under the control of the operating system. At this time, the processor can execute commands corresponding to the program code loaded into memory for each of the patient expression acquisition unit (110) and the drug recommendation unit (120), thereby executing subsequent steps (210 to 220), which may be different functional representations of the processor.

[0023] In step (210), the patient expression acquisition unit (110) can acquire a patient expression representing the patient's current health status by using the patient's current visit and the patient's past visits, including the diagnosis and surgical history information prescribed. Here, a visit may mean a patient's visit to a medical institution.

[0024] In step (220), the drug recommendation unit (120) may recommend a drug to the patient based on the acquired patient expression, using first drug history information reflecting the degree of prescription between drugs, past health states of patients similar to the patient's current health state, and second drug history information prescribed based on the past health states of other patients. The drug recommendation unit (120) may acquire a drug expression by considering the relationship information of co-prescription between drugs and the relationship information of low possibility of causing side effects when drugs are taken together through the EHR graph+ and the DDI (Drug Drug Interaction) graph. At this time, the EHR graph+ may be an EHR graph in which weight direction is considered. The drug recommendation unit (120) may use the EHR graph+ to represent all drugs, and two drugs of each drug pair, as directed edges, and may assign the degree of co-prescription between drugs considered at each drug level as a weight of the edge. The medication recommendation unit (120) can obtain a first medication recommendation expression required for the patient's current visit based on the acquired medication expression and the acquired patient expression. The medication recommendation unit (120) can obtain a second medication recommendation expression required for the patient's current visit by reflecting the degree to which the patient expression of the past visit in which each medication was prescribed and the patient expression of the current visit are related based on the patient expression obtained through health state-aware attention and medications prescribed in past visits that show other patient expressions similar to the acquired patient expression. The medication recommendation unit (120) can derive medications required for the patient's current visit by merging the first medication recommendation expression and the second medication recommendation expression.

[0025] FIG. 3 is a diagram for explaining a drug recommendation operation in one embodiment.

[0026] The drug recommendation system is based on the patient's first visit (V 1 ) to visit T (T is a natural number) (VT ) and the diagnosis (d) / surgery (p) at the patient's first visit (V 1 ) from the T-1 visit (V T-1 ) can extract accurate medications for the patient's T visit through a medication recommendation model. In other words, the medication recommendation system ... based on the patient's health record. and the diagnosis and surgical history prescribed at the current visit T d T , p T , and given Health Status-Aware Attention and EHR Graph+, it can recommend the most accurate medications at the patient's current visit T.

[0027] The drug recommendation system uses the patient representation q, which represents the patient's current health status, to utilize the patient's diagnosis and surgical history prescribed in the patient's current and past visits through an encoder. T At this time, the drug recommendation system selects only past visits related to the present among all past visits of the patient based on the diagnosis history and surgical history, and then calculates the degree of relevance between the present and the selected past visits. By calculating and weighting the current and selected past visits based on the calculated relevance, a patient representation, which is an improved representation of the patient's health status, can be obtained.

[0028] The drug recommendation system uses the patient's current visit T and past visits t, and the diagnosis and surgical history d of each visit. T , p T Concatenate the weight matrix W of the embedding layer e By obtaining the weight matrix of the obtained embedding layer and feeding it to the embedding layer, we obtain a dense representation v for the current visit and all past visits. T , vt can be obtained (v t =concat(d T , p T )W e ). The drug recommendation system uses a dense representation of the patient's current visit and each past visit v T , v t Concatenate the weight matrix W of the MLP layer S By obtaining the weight matrix of the obtained MLP layer and feeding it to the MLP (Multi-Layer Perceptron) layer, the probability s that each past visit t is selected as a past visit related to the current visit t is obtained. t can be obtained(s t =sigmoid(concat(v T , v t )W S +b s , b s represents the bias value of the MLP layer). The drug recommendation system is based on the probability s that each past visit t will be selected. t and the probability of not being selected (1-s) t ) into a two-dimensional vector, and then apply Gumbel-softmax to these vectors to determine whether to select a relevant past visit for each past visit t. The drug recommendation system, given the past visits related to the patient's current (i.e., previously selected) visits, can create a dense representation v of the patient's current visit. T query, all dense representations v for past visits t related to the current visit T T and v t The degree of correlation between the patient's current visit T and past visits t using an attention network that uses as keys and values can be calculated. The drug recommendation system can compute all dense representations v for past visits t related to the current visit T. T , v t Degree of relevance to Patient representation q, which represents the patient's current health status by weighting the sum of T can be obtained.

[0029] The drug recommendation system can recommend accurate and safe medications to patients by utilizing the patient representation acquired from the encoder via a predictor, the Health Status-Aware Attention and EHR Graph+ (HSTAT) graph, and the Drug Drug Interaction (DDI) graph, which indicates the relationship between side effects among medications. The drug recommendation system can use the same EHR Graph+ graph for all patients.

[0030] The medication recommendation system utilizes EHR Graph+ and DDI (Drug Drug Interaction) graphs to generate medication representation e for all medications i. i can be obtained. The drug recommendation system is based on the patient expression and the medication representation. and second drug expression At this time, the drug recommendation system can obtain drug expressions while considering the relationship between drugs being prescribed together and the relationship that is less likely to cause side effects when taken together.

[0031] The drug recommendation system considers the patient representation based on the drug representation and the patient representation. The first drug representation (current health-aware medication feature representation) captures the features of the kth drug required in the patient's current visit, and the health status-aware attention is utilized to select a patient representation similar to the current patient representation without distinguishing between the patient's own past visits and the visits of other patients. For medications prescribed in past visits, a second medication representation (current health-relevant past medication representation) can be obtained that reflects how closely the patient's representation in the past visits in which each medication was prescribed is related to the current patient representation. The medication recommendation system can then fuse the first and second medication recommendation representations to ultimately derive the medications required for the patient's current visit.

[0032] The drug recommendation system refines the existing EHR graph into EHR Graph+ by representing all drug pairs as directed edges from the perspective of each drug, and assigning a weight to each edge based on the degree of co-prescription considered at each drug level. At this time, elements i and j belonging to the set of directed edges between drugs indicate that drugs i and j have been prescribed together for any patient in any visit, and the weight is established only if drugs i and j have been prescribed together for any patient in any visit, and has the value of ((the number of visits in which drugs i and j were prescribed together) / the total number of visits in which drugs i was prescribed)) (the same applies to j and i).

[0033] More specifically, the drug recommendation system can effectively reflect the degree of co-prescription considered at each drug level by representing all drug pairs through EHR Graph+, with the two drugs in each drug pair represented by an edge with a direction pointing toward the other from the perspective of each drug, and assigning a weight to each edge based on the degree of co-prescription considered at each drug level. EHR Graph+ can be defined as follows: where, represents a set of edges with directionality between drugs, Elements belonging to (i, j)∈ indicates that two drugs i and j were prescribed together for a patient in a certain visit, and (i, j)≠(j, i). Therefore, the adjacency matrix satisfies the following conditions: = (Number of visits where drugs i and j were prescribed together) / (Total number of visits where drug i was prescribed). Therefore, if drugs i and j were prescribed together in any visit for any patient, >0; otherwise =0.

[0034] The drug recommendation system uses EHR graph+ and DDI graph to randomly initialize drug representation e i After applying two-layer DiGCN and two-layer GCN respectively, the outputs of the two GCNs for each drug can be fused through a subtraction operation. The drug recommendation system can be used to represent drugs e iGiven this, by leveraging the patient representation obtained from the encoder along with the drug representation, we can obtain a first drug representation that captures the characteristics of the kth drug required for the patient's current visit T, given the patient representation. The drug recommendation system can then use previously predicted drugs to identify the characteristics of drugs that are required but not yet predicted, given the patient representation. At this point, the drug recommendation system can also consider the relationships between the predicted drugs. This process can be formally expressed as follows.

[0035]

[0036] Here, refers to a transformer-based model, represents the k-1th predicted drug expression in the patient's current visit, represents a dot-product. This first drug expression can also be utilized in the process of obtaining a second drug expression.

[0037] A drug recommendation system can obtain a secondary drug representation that reflects the degree to which similar patient expressions prescribed in past visits with similar patient expressions are related to the patient expression, regardless of patient type. This relatedness indicates the likelihood that each drug will be recommended as the kth drug needed in the patient's current visit. The drug recommendation system can obtain a secondary drug representation that captures the likelihood that each of these drugs, prescribed in past visits with similar patient expressions, will be recommended as the kth drug needed in the patient's current visit, regardless of patient type.

[0038] In obtaining a second drug representation, the drug recommendation system can flexibly select only past visits that show a health condition similar to the patient's current visit health condition (patient representation), regardless of patient classification, and then utilize the drug history prescribed in the selected past visits.

[0039] Here, health-state-aware attention pre-selects patient representations, which represent the patient's health status. More specifically, a drug recommendation model is trained on a training set, and then hyperparameters are tuned using a validation set. After tuning, patient representations for all visits of all patients are obtained from the finalized drug recommendation model. Based on the obtained patient representations, cosine similarity is used to pre-select the top N similar past visits (for all patients' current visits).

[0040] The drug recommendation system is a set of top-N visits that are flexibly selected without distinction between patients with similar health status to the patient's current health status in health-state-aware attention. ( and ) can be obtained. The drug recommendation system can obtain the patient's current visit T and the health-state-aware attention. Information related to past visits included in can be obtained at two levels (medical level and visit level). Degree of relevance at medical level Is Indicates how each drug i prescribed in a past visit is related to the kth drug feature (e.g., the first drug expression) required in the patient's current visit.

[0041]

[0042] Here, represents the drug expression of drug i prescribed in the past visit. Meanwhile, the degree of relevance at the visit level Is Indicates how the health status of the patient's current visit is related to the health status of the patient's past visits included in the drug recommendation system. Similar patient representations from past visits included in the dataset can be obtained from the encoder. Then, the drug recommendation system uses the patient representation as a query. The degree of relevance at the visit level is determined by an attention network that uses similar patient representations for all past visits as keys and values. can be calculated.

[0043]

[0044] Here, represents the weight matrix of the attention network.

[0045] The drug recommendation system is based on the drug level Wow visit level After obtaining the second drug representation in health-state recognition attention through fusion can be obtained.

[0046]

[0047] Finally, the drug recommendation system can fuse the first drug representation and the second drug representation to obtain the kth drug required for the patient's current visit.

[0048]

[0049] Here, For all medications, represents the score that each medication should be predicted to be the kth medication needed at the patient's current visit. represents a learnable parameter that controls the balance of two drug expressions. class represent the weight matrix and bias vector, respectively.

[0050] Drug representation e by utilizing cross entropy loss for drugs predicted by the drug recommendation system i and other learnable parameters can be learned.

[0051]

[0052] Here, represents 1 if drug i was (actually) prescribed at visit t, and 0 otherwise.

[0053] In an embodiment, an experiment can be conducted to derive the technical effectiveness of a drug recommendation method. MIMIC-III, a publicly available dataset from US intensive care units, was used as the experimental data. Only data from patients who were hospitalized more than twice were used. The training / validation / test data split ratio can be set to 4:1:1. The representation dimension can be 64, the batch size 16, the epoch 100, and the learning rate 0.001 can be identically set. The Gumbel softmax parameter can be fixed to 0.8, the attention network parameter 20, and the number of visits drawn can be fixed to 3. The effectiveness of the proposed method can be verified through extensive experiments using real-world MIMIC-III and MIMIC-IV datasets. Most importantly, the proposed method is more accurate than the best competitor, MoleRec, by 18.69% and 17.26% in terms of Jaccard criterion, respectively, while recommending comparable safe drugs. These experimental results confirm that the method proposed in the examples improves accuracy while recommending safe drugs within the range of acceptable actual prescriptions.

[0054] The devices described above may be implemented as hardware components, software components, and / or a combination of hardware components and software components. For example, the devices and components described in the embodiments may be implemented using one or more general-purpose computers or special-purpose computers, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), a programmable logic unit (PLU), a microprocessor, or any other device capable of executing instructions and responding to them. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to the execution of the software. For ease of understanding, the processing device is sometimes described as being used alone; however, one of ordinary skill in the art will recognize that the processing device may include multiple processing elements and / or multiple types of processing elements. For example, a processing unit may include multiple processors, or a processor and a controller. Other processing configurations, such as parallel processors, are also possible.

[0055] Software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing device to perform a desired operation or may independently or collectively command the processing device. The software and / or data may be embodied in any type of machine, component, physical device, virtual equipment, computer storage medium, or device for interpretation by the processing device or for providing instructions or data to the processing device. The software may also be distributed over networked computer systems and stored or executed in a distributed manner. The software and data may be stored on one or more computer-readable recording media.

[0056] The method according to the embodiment may be implemented in the form of program commands that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program commands, data files, data structures, etc., alone or in combination. The program commands recorded on the medium may be those specially designed and configured for the embodiment or may be those known and available to those skilled in the art of computer software. Examples of the computer-readable recording medium include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, and hardware devices specially configured to store and execute program commands such as ROMs, RAMs, and flash memories. Examples of program commands include not only machine language codes generated by a compiler, but also high-level language codes that can be executed by a computer using an interpreter, etc.

[0057] Although the embodiments described above have been described by way of limited examples and drawings, those skilled in the art will appreciate that various modifications and variations can be made based on the above teachings. For example, appropriate results can still be achieved even if the described techniques are performed in a different order than described, and / or components of the described systems, structures, devices, circuits, etc. are combined or combined in a different manner than described, or are replaced or substituted with other components or equivalents.

[0058] Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims described below.

Claims

1. In a drug recommendation method performed by a drug recommendation system, A step of obtaining a patient representation representing the patient's current health status by using the patient's current visit and the patient's diagnosis history information and surgical history information prescribed in the patient's past visits; and A step of recommending a medicine to a patient by using first medicine history information reflecting the degree of prescription among medicines based on the patient expression obtained above and second medicine history information prescribed based on the past health status of a patient similar to the current health status of the patient and the past health status of another patient. A method of recommending medicines including:

2. In paragraph 1, The above recommended steps are: A step for obtaining a drug representation by considering the relationship information between drugs prescribed together and the relationship information that is less likely to cause side effects when taken together, through EHR graph+ and DDI (Drug Drug Interaction) graph. A method of recommending medicines including:

3. In paragraph 2, The above recommended steps are: Using EHR Graph+, a step is taken to represent all medicines as directed edges for each pair of medicines, and to assign weights to the edges based on the degree to which the medicines are prescribed together at each medicine level. A method of recommending medicines including:

4. In paragraph 2, The above recommended steps are: A step of obtaining the first drug expression required for the patient's current visit based on the obtained drug expression and the obtained patient expression. A method of recommending medicines including:

5. In paragraph 1, The above recommended steps are: A step of obtaining a second medication expression required for the patient's current visit by reflecting the degree to which the patient's expression in the past visits in which each medication was prescribed is related to the patient's expression in the current visit based on the patient expression obtained above and other medications prescribed in past visits showing other patient expressions similar to the obtained patient expression through Health Status-Aware Attention. A method of recommending medicines including:

6. In paragraph 1, The above recommended steps are: A step to derive the medicines required for the patient's current visit by merging the first and second medicine expressions. A method of recommending medicines including:

7. In paragraph 5, The above Health Status-Aware Attention is a drug recommendation method characterized in that, after learning a drug recommendation model using a learning set, hyperparameter tuning of the learned drug recommendation model is performed using a verification set, patient expressions for all visits of all patients are obtained from the drug recommendation model confirmed through the performed hyperparameter tuning, and a specific number of visits having a preset rank or higher are selected from among similar past visits based on cosine similarity based on the obtained patient expressions.

8. In the drug recommendation system, A patient expression acquisition unit that acquires a patient expression representing the patient's current health status by using the patient's current visit and the patient's diagnosis history information and surgical history information prescribed in the patient's past visits; and A drug recommendation unit that recommends drugs to a patient by using first drug history information reflecting the degree of prescription among drugs based on the patient expression obtained above, and second drug history information prescribed based on the past health status of a patient similar to the current health status of the patient and the past health status of another patient. A drug recommendation system including:

Images