CN116312915B - A method and system for standardized association of drug terms in electronic medical records - Google Patents

Abstract

The invention discloses a method and a system for standardized association of drug terms in electronic medical records, wherein a drug term library is updated through a synonym mining technology to obtain a drug term library based on synonym mining update, so that the problem of low semantic similarity between standard drug terms in the drug term library and external drug terms in the electronic medical records is solved; when the external medicine terms in the electronic medical record are associated with standard medicine terms in the medicine term library based on synonym mining update, the semantic information is utilized to add pinyin character sequences of corresponding medicine terms except Chinese character marks, and the diagram structure information of the external medicine terms in the medicine term library and the electronic medical record is fully utilized; an association prediction model based on semantic embedding and structural embedding is constructed, so that the association between external medicine terms and standard medicine terms in a medicine term library in the real-world electronic medical record is accurately established.

Description

A method and system for standardized association of drug terms in electronic medical records

Technical field

The invention belongs to the field of medical information technology, and in particular relates to a method and system for standardized association of drug terms in electronic medical records.

Background technique

With the development of information technology and its continuous in-depth application in the medical and health industry, a large amount of data is stored in the medical and health industry. Among them, the more typical ones include knowledge base (KB) presented in a relatively standardized form and electronic health records (EHRs) presented in the form of real-world diagnosis and treatment process data. Knowledge base is a technology used by computer systems to store complex structured and unstructured information. Terminology base (TB), as a special type of knowledge base, is used to store terminology concepts and related information. In the field of drug research, general drug terminology databases have been established and are still being continuously updated and improved, including Drugbank and WHODrug. In addition, in academia and industry, there are also Chinese drug terminology databases established in response to demand. However, in the practice of real-world clinical diagnosis and treatment, different regions and even different hospitals and different doctors may use multiple different names to refer to the same drug, and existing drug terminology databases may not be able to record all names of drugs. For example, the English name of the drug with Drugbank ID=DB00736 is "Esomeprazole magnesium", the Chinese name was "Esomeprazole Magnesium", and the current name is "Esomeprazole Magnesium", and in Electronic Health In the medical record system, it may be recorded as "esomeprazole magnesium" before the drug's generic name is modified, and it may be recorded as "esomeprazole magnesium" after the drug's generic name is modified. When using electronic health record data to conduct real-world During drug research, if any of these names are missed, incomplete data retrieval will lead to unreasonable screening of the research population and incorrect estimation of drug use, ultimately affecting the quality of the research. Therefore, when using EHRs data to conduct real-world drug research, especially multi-center real-world drug research involving multiple drugs, it is very necessary to associate the drug names in the EHRs with the corresponding drugs in the drug glossary. An important prerequisite to ensure the quality of research and the reliability of results. As important information in the field of medical research and engineering, drug terminology databases are updated in a timely manner and are the basis for promoting information exchange and even technological progress in this field. Correlating it with real-world electronic health record data can provide a basis for natural language processing based on EHRs, Research and engineering tasks in artificial intelligence, expert systems, real-world drug research, etc. provide underlying support and promote and promote.

Among the existing drug association methods, a universal model-based medical standard terminology management system and method (publication number CN115080751A) involves the mapping of medical record text and standard terminology. First, the medical record text is split based on the sequence annotation model to obtain text details. attributes, and then calculate the similarity with any semantic standard word, and judge the effectiveness of the standardized mapping through the semantic similarity. If it is valid, it will be directly used as the mapping result. If it is invalid, other possible standardized mappings will be recalculated, and finally used as an algorithm. Recommended mapping results that require manual review. However, this technical solution only uses semantic similarity to judge the validity of the mapping and ignores the structural characteristics of the drug term database.

A drug name matching method and device (publication number CN112711642A) involves drug matching between different electronic medical records. First, the word vector of the electronic medical record corpus is obtained through electronic medical record data training, and the drug name is extracted based on the unified medical language system. Obtain the drug entity word vector, and use the neural network model to obtain the component vector. At the same time, combined with engineering features, the similarity between drug entities is calculated, and finally drug matching between different electronic medical record systems is achieved. This technical solution solves the problem of drug matching between different electronic health medical record data under the assumption that the unified medical language system is perfect. The problem to be solved by the present invention is to match drug terms in the electronic health medical record to the drug term library. The reference value is limited.

The drug information matching method and system (publication number 107103048B) involves matching between drugs. First, multiple dimensions of sub-information of the drugs to be matched are obtained, such as drug names, preparation specifications, dosage forms, etc., and the target sub-information is compared with the standard sub-information. The information is identified for correlation. When the correlation identification result meets the preset correlation requirements, the target information that meets the preset correlation requirements and one or more standard information are respectively configured into one or more candidate information pairs; for each candidate information pair , calculate the similarity between the target information and the standard information on the sub-information of multiple dimensions, and calculate the comprehensive matching score of each candidate information pair based on the calculated similarity. Finally, the maximum comprehensive matching score corresponding to The standard information in the candidate information pair is determined as the matching information of the target information. The medical drug matching method, device, electronic equipment and storage medium (publication number is CN111798969A) involves the matching between the target drug and the drug standard library. The method of the application includes: for the target drug to be matched, select from drug information Multiple drug labels or specifications used to characterize the target drug are used as benchmark items, and a weight value is assigned to each benchmark item based on importance. The benchmark items are matched with the standard items in the drug benchmark library, and the comparison value is calculated. Based on the comparison value and weight value, the matching degree of the target drug and the drugs in the drug benchmark library is calculated, thereby establishing a mapping relationship between the drug identifier of the target drug and the standard identifier of the target drug stored in the drug benchmark library. These two technical solutions solve the matching problem of target drugs and drug standard libraries. Compared with drugs, drugs contain more sub-information. In addition to drug names, they also include preparation specifications, dosage forms, manufacturers, approval numbers, etc. However, since the problem to be solved by the present invention is drug association, the text information that can be used is limited, so these methods are not applicable.

The limitations of the existing technology are mainly reflected in: in the association process, only semantic similarity is used, but the graph structure information is not used; in the association process, semantic similarity does not use pinyin information, because the drug names may have different words but If the pronunciation is the same, if you simply use the semantic similarity of the Chinese name, for example, "cefradine" may be calculated as having close similarity with "cefradine" and "ceftazidime", but from the pinyin point of view, it is obvious that "cefradine" "Cefradine" and "cefradine" are the same drugs, resulting in inaccurate correlation results.

Contents of the invention

The purpose of the present invention is to provide a standardized association method and system for drug terms in electronic medical records in view of the shortcomings of the existing technology, so as to realize the association between external drug terms in electronic medical records and standard drug terms in the drug term library.

The purpose of the present invention is achieved through the following technical solutions:

According to the first aspect of this specification, a standardized association method for drug terms in electronic medical records is provided, including:

S1, input the drug term database and obtain the synonym set of each standard drug term;

S2, obtain an updated drug terminology database based on synonym mining, including:

Constructing a corpus for synonym mining and obtaining a list of drug terms from said corpus;

Train a synonym set classifier to obtain classification prediction results for each drug term in the drug term list and the synonym set in the drug term database, and obtain all synonym sets updated based on synonym mining according to a preset probability threshold;

Update the drug terminology database according to all synonym sets updated based on synonym mining;

S3, based on the updated drug term database and external drug terms in the electronic medical record, train an association prediction model based on semantic embedding and structural embedding, including:

The semantic embedding representation of the pairs of external drug terms in the electronic medical record and the standard drug terms in the updated drug term database is obtained through the pre-trained language model, specifically: the external drug terms and their pinyin character sequences, the standard drug terms and The pinyin character sequence, combined with the starting character and the separator character, forms a character sequence of associated drug term pairs, which is input into the pre-trained language model to obtain a semantic embedding representation;

Obtain the structural embedded representation of the pair of external drug terms in the electronic medical record and the standard drug terminology in the updated drug terminology database through the graph convolutional neural network model, specifically: combining the external drug terminology with the updated drug terminology database Drug terms establish candidate association relationships based on similarity calculations, use the semantic embedding representations of the external drug terms and the drug terms in the updated drug term database as initialization node embedding representations of the corresponding drug terms, and input the graph convolutional neural network Model, obtain the node embedding representation of the corresponding drug term, and use the product of the node embedding representation of the external drug term and the standard drug term as the structure embedding representation;

S4: Use the association prediction model to predict the association results between external drug terms in the electronic medical record and standard drug terms in the drug term database.

Further, during the training process of the synonym set classifier, the probability that the drug term to be classified belongs to the synonym set is predicted based on the set unity score change. The set unity score calculation method is: calculate the embedded representation of each term in the set , input the embedding representation into the fully connected neural network model to obtain a new term representation, add all new term representations and take the average to obtain the initialized term set representation, input the initialized term set representation into the fully connected neural network model to obtain the set Unity score.

Further, the training set generation method of the synonym set classifier includes: extracting a drug term from the synonym set by random extraction, and generating a positive training sample by combining the set of other drug terms in the synonym set; for each positive training sample, Match multiple negative training samples, which are extracted from the drug term database after excluding drug terms in the synonym set.

Further, the drug term database is updated according to all the synonym sets updated based on synonym mining, specifically: if the synonym set of the standard drug term as a hypernym is updated, the corresponding synonyms are compared with the standard drug as a hypernym. Establish synonym associations for the terms, and at the same time establish hyponym associations between the corresponding synonyms and all standard drug terms associated with the standard drug terms as hypernyms; if the synonym set of non-hypernym standard drug terms is updated, associate the corresponding synonyms with The non-hypernym standard pharmaceutical terminology establishes synonym associations.

Further, the pre-trained language model is adjusted, specifically: the semantic embedding representation of the starting character is used as the independent variable, and the dependent variable is whether the external drug terms in the electronic medical record and the drug terms in the updated drug term database are semantically related or not. Label; use a nonlinear activation function to obtain prediction results based on semantic embedding representation; use manual annotation to obtain positive training samples, and use random extraction to obtain negative training samples to obtain a training set, perform pre-training language model training, and optimize and adjust the semantic embedding representation. .

Further, a candidate association relationship is established between the external drug term and the drug term in the updated drug term database, specifically: calculating the TF-IDF value of each word in the drug term, and obtaining the relationship between the external drug term in the electronic medical record and The vector representation of each drug term in the updated drug term database is calculated, and the similarity between the two drug terms is calculated. If the similarity is greater than the preset similarity threshold, the external drug term in the electronic medical record is compared with the corresponding drug. Candidate association relationships are established for drug terms in the term database.

Further, the input of each layer in the graph convolutional neural network model includes two parts. The first part is the node embedding representation matrix, and the second part is the adjacency matrix. The output of each layer is used as the node embedding representation matrix of the next layer. Obtained through the normalized graph Laplacian transform, a marginal-based distance loss function is used to optimize the graph convolutional neural network model.

Further, the value of the adjacency matrix is specifically: if there is an edge from one drug term to another drug term in the updated drug term database, then the corresponding value is 1, otherwise the value is 0; if there is An edge from an external drug term in the electronic medical record to a drug term in the updated drug term database, then the corresponding value is the similarity value in the candidate association relationship.

Further, the semantic embedding representation and the structure embedding representation are spliced, and the spliced representation is input into a multi-layer perceptron. The multi-layer perceptron includes multiple fully connected hidden layers and a single-node output layer. The multi-layer perceptron is input into the multi-layer perceptron. The output of the layer perceptron is converted into a scalar through a nonlinear activation function to obtain the association probability of external drug terms in each electronic medical record with standard drug terms in the updated drug term database.

According to the second aspect of this specification, a standardized association system for drug terms in electronic medical records is provided, including:

The electronic medical record drug term input module is used to obtain all external drug terms to be standardized in the electronic medical record;

The drug term database synonym mining update module is used to construct a corpus, obtain a drug term list for synonym mining from the corpus; obtain a synonym set for each standard drug term in the drug term database; train a synonym set classifier to obtain all Classification prediction results of each drug term in the drug term list and the synonym set in the drug term database are obtained, and all synonym sets updated based on synonym mining are obtained according to the preset probability threshold, and the drug term database is updated;

The candidate association relationship establishment module is used to establish a candidate association relationship between external drug terms in the electronic medical record and drug terms in the updated drug term database based on similarity calculation;

The semantic embedding representation module is used to combine external drug terms and their pinyin character sequences in electronic medical records, standard drug terms and their pinyin character sequences in the updated drug term database, combined with starting characters and delimiter characters to form an associated drug term pair character sequence. , input the pre-trained language model and obtain the semantic embedding representation;

The structural embedding representation module is used to use the semantic embedding representation of external drug terms in the electronic medical record and the updated drug terminology database as the initialization node embedding representation of the corresponding drug terms, and input the graph convolutional neural network model to obtain the corresponding drug. The node embedding representation of the term is a product of the node embedding representation of the external drug term and the standard drug term as a structural embedding representation;

The association prediction module is used to train an association prediction model based on semantic embedding and structural embedding based on the updated drug term database and external drug terms in the electronic medical record; use the association prediction model to predict and obtain the external drug terms and drug terms in the electronic medical record Association results for standard drug terms in the library.

The beneficial effects of the present invention are: the present invention enriches semantic information and graph structure information through synonym mining technology; in association prediction model prediction, semantic information and graph structure information are used simultaneously; specifically embodied in:

1) Update the drug terminology database through synonym mining technology to obtain an updated drug terminology database based on synonym mining, which solves the problem of low semantic similarity between standard drug terms in the drug terminology database and external drug terms in electronic medical records;

2) When the external drug terms in the electronic medical record are associated with the standard drug terms in the drug term database updated based on synonym mining, the semantic information used is to add the pinyin mark of the corresponding term in addition to the Chinese character mark;

3) Make full use of the graph structure information of the drug terminology database when associating external drug terms in the electronic medical record with standard drug terms in the drug terminology database updated based on synonym mining;

4) When the external drug terms in the electronic medical record are associated with the standard drug terms in the drug term database updated based on synonym mining, obtain the external drug terms in the electronic medical record by associating the external drug terms in the electronic medical record with the drug terms in the drug term database. graph structure information;

5) Through the above method, the embedded representation information of the external drug terms in the electronic medical record and the standard drug term pair in the drug term database updated based on synonym mining is finally obtained, and the prediction of the association prediction model is performed.

Description of drawings

Figure 1 is an overall step flow chart of a standardized association method for drug terms in electronic medical records provided by an exemplary embodiment;

Figure 2 is a flow chart for implementing a method for standardized association of drug terms in electronic medical records provided by an exemplary embodiment;

Figure 3 is a schematic diagram of the original drug term database provided by an exemplary embodiment;

Figure 4 is a schematic diagram of a drug term database updated based on synonym mining provided by an exemplary embodiment;

Figure 5 is a structural diagram of a standardized association system for drug terms in electronic medical records provided by an exemplary embodiment.

Detailed ways

In order to make the above objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

Many specific details are set forth in the following description to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described here. Those skilled in the art can do so without departing from the connotation of the present invention. Similar generalizations are made, and therefore the present invention is not limited to the specific embodiments disclosed below.

As shown in Figures 1 and 2, an embodiment of the present invention provides a method for standardized association of drug terms in electronic medical records, including the following steps:

Step S1: Enter the drug term database, which is expressed as , where E represents a set of drug terms, including standard drug terms as hypernyms or non-hypernyms, and synonyms of standard drug terms, and R represents a set of relationships between drug terms, specifically, /> , of which/> Indicates a hyponym relationship,/> Representing synonym relationships, the relationship between drug terms can be expressed as , representing that the drug term h is a synonym for t, or , indicating that the drug term h is a hyponym of t; the relationship in the drug term database is/> Convert drug terms into synonym sets to obtain synonym sets for each standard drug term;

Figure 3 is an example of a drug term database, in which standard drug terms include esomeprazole, esomeprazole strontium, esomeprazole strontium hydrate, esomeprazole sodium, esomeprazole magnesium, Esomeprazole magnesium dihydrate, esomeprazole magnesium trihydrate, gadopentetate, gadopentetate dimeglumine, laoxycephalosporin, laoxycephalosporin sodium; the hyponyms of esomeprazole include Esomeprazole strontium, esomeprazole strontium hydrate, esomeprazole sodium, esomeprazole magnesium, esomeprazole magnesium dihydrate and esomeprazole magnesium trihydrate, Taking esomeprazole strontium as an example, the relationship can be expressed as ; Synonyms of laoxycephalosporin include latacephalosporin, and hyponyms include laoxycephalosporin sodium; hyponyms of gadopentetate include gadopentetate dimeglumine, and synonyms of gadopentetate dimeglumine include gadopentetate dimeglumine The relationship between amine and gadopentetate dimeglumine, gadopentetate dimeglumine and gadopentetate dimeglumine can be expressed as ;Standard drug terminology is the target association object of this embodiment.

Step S2: In order to solve the problem of incomplete synonym relationships in the original drug terminology database due to problems such as irregular translation, inaccurate manual annotation, and untimely data updates, the synonym mining method is used to improve the synonym relationships of the drug term database and obtain updated synonyms based on synonym mining. Drug term database; specifically includes the following sub-steps:

Step S21: Obtain Chinese abstracts and text of drug-related Chinese literature from literature retrieval platforms such as CNKI and Wanfang to form a corpus for synonym mining , and use named entity recognition method to obtain drug terms for synonym mining, expressed as a drug term list/> , of which/> Represents the i-th drug term,/> Express/> The number of Chinese medicine terms, the named entity recognition method used in this embodiment is a conditional random field model;

Note all synonym sets in the drug glossary as , of which/> Represents the synonym set of the jth standard drug term,/> Express/> The number of synonym sets in /> It is equal to the number of all standard drug terms in the drug term database. If a standard drug term has no synonyms, its corresponding synonym set only contains 1 element, which is the standard drug term.

Step S22: Train the synonym set classifier to obtain a list of drug terms The classification prediction results of each drug term in the drug term database and the synonym set in the drug term database; specifically include the following sub-steps:

Step S221: Express the synonym set classifier as , of which/> Represents a synonym set,/> Represents a drug term to be classified into a synonym set;

Step S222: Predict drug terms to be classified based on changes in set unity scores Belongs to synonym set/> The probability of , the formula can be expressed as/> , where Pr represents probability, is the sigmoid activation function,/> is the set unity score function;

Step S223: For a set of data , of which/> Represents a synonym set/> and to be classified into synonym sets/> Drug terminology in /> ,/> Indicates label,/> Express/> ,/> Express/> ,/> , ; In this embodiment, the synonym set classifier adopts a fully connected neural network model, and the loss function of the synonym set classifier adopts a logarithmic loss function, which is specifically expressed in the following form:

.

In one embodiment, the set unity score Estimated by:

First for the term set Every term in/> , use text embedding method to calculate its embedding representation/> , as the initialization parameter of the embedding layer, input the embedding layer. The text embedding method used in this embodiment is Word2Vec; then enter the embedding representation into the fully connected neural network model to obtain a new term representation of the corresponding term/> , convert the term set/> Corresponding to all new terminology/> After performing the addition operation, the average value is taken to ensure the invariance of substitution and the initialized term set representation is obtained, that is,/> ;Finally represent the above term set/> Enter the fully connected neural network model to get the final term set/> unity score/> , used to measure the set/> The degree of similarity of all terms within.

In one embodiment, the training set is generated by:

First, the relationship in the drug term database is Drug terms are converted into synonym sets, and all synonym sets in the drug term database are expressed as/> ;Each synonym set is expressed as/> , of which/> Represents each drug term in the synonym set; randomly selects any drug term from the set ES/> , the remaining terms in the set ES constitute the set , get the positive samples used for model training/> , label y=1; for each positive sample, match K negative samples, expressed as/> , label y=0, in this example, K is 5, where/> It can be obtained by extracting from the drug term database after excluding the drug terms in the synonym set ES. Specifically, it can be obtained by mixing samples extracted in the following two ways according to a set proportion: ① Using completely random sampling. Obtained by extraction; ②Limit the sampling range to the set/> The drug terms in contain drug terms with the same characters, and are extracted by random sampling; in this embodiment, the ratio is set to 2:3.

Step S224: Use the training set to train the model, and predict the drug term list based on the trained model The probability that each drug term in belongs to each synonym set in the drug term base;

Specifically, before synonym mining is performed, all synonym sets in the drug term database are expressed as ;for/> Every drug term in /> , calculate which belongs to/> Any synonym set in/> The probability of , expressed as/> , take the maximum probability value as the drug term/> Potential classification, set probability threshold/> , if the above probability maximum value is greater than the above probability threshold/> , the corresponding synonym set will be updated, if the above probability maximum value is less than or equal to the above probability threshold/> , change the above drug terms/> Return to/> ;Start the next cycle until The probability that all drug terms in belong to any synonym set is less than or equal to the above probability threshold/> , and finally obtain all synonym sets updated based on synonym mining, expressed as/> ; In this embodiment, take/> .

Step S23: Update the drug terminology database based on all the synonym sets updated based on synonym mining, and obtain a drug term database updated based on synonym mining;

Specifically, if the synonym set of the standard drug term as a hypernym is updated, in addition to establishing a synonym association between the corresponding synonym and the standard drug term as a hypernym

In addition, at the same time, establish a hyponym association between the corresponding synonyms and all non-hypernym standard drug terms associated with the standard drug term as a hypernym. The relationship is expressed as

. If the synonym set of a non-hypernym standard drug term is updated, establish a synonym relationship between the corresponding synonym and the non-hypernym standard drug term. The relationship is expressed as

. In the example in Figure 4, the updated synonym set based on synonym mining: [{esomeprazole, esomeprazole}, {esomeprazole sodium, esomeprazole sodium},…], Establish a synonym relationship between the synonym esomeprazole and the standard drug term esomeprazole as a hypernym, and relate the synonym esomeprazole to the standard drug term esomeprazole as a hypernym. Standard drug terminology establishes hyponym associations: esomeprazole strontium, esomeprazole strontium hydrate, esomeprazole sodium, esomeprazole magnesium, esomeprazole magnesium dihydrate, Esso Meprazole magnesium trihydrate; establish a synonym association between esomeprazole sodium and esomeprazole sodium; and finally obtain an updated drug term database based on synonym mining.

Step S3: Based on the above-mentioned updated drug terminology database based on synonym mining and external drug terms in real-world electronic medical record data, train an association prediction model based on semantic embedding and structural embedding; specifically including the following sub-steps:

Step S31: Obtain semantic embedding representations of external drug terms in the electronic medical record and standard drug term pairs in the drug term database updated based on synonym mining through a pre-trained language model. In this embodiment, the pre-trained language model adopts the BERT model;

Specifically, the set of external drug terms in real-world electronic medical records is represented as , the set of standard drug terms in the drug term database updated based on synonym mining is expressed as/> , any external drug term in the external drug term set G in the real-world electronic medical record /> The pinyin character sequence is expressed as , any standard drug term in the standard drug term set E in the updated drug term database based on synonym mining/> The Pinyin character sequence is expressed as/> , will/> Combine the starting character [CLS] and the delimiting character [SEP] to form a character sequence of associated drug term pairs, recorded as /> , with/> For "esomeprazole sodium", /> Taking "esomeprazole sodium" as an example, the character sequence of the associated drug term can be expressed as {[CLS][Ai][Division][Ao][Me][L][Azole][Sodium][ai][ si][ao][mei][la][zuo][na][SEP][Angstrom][Soul][American][La][azole][Sodium] [ai][suo] [mei][la] [zuo][na] [SEP]}; In this embodiment, the BERT model pre-trained based on the Chinese corpus is used, and through multiple bidirectional encoding layers of the Transformer, the semantic embedding representation of the character sequence of the above-mentioned associated drug terms is obtained, and finally used represented by the semantic embedding representation of the initial character [CLS]/> and/> correlation.

In this embodiment, the BERT model is adjusted, specifically: the semantic embedding representation of the starting character is used as an independent variable, recorded as , the dependent variable is external drug terms in real-world electronic medical records/> and standard drug terms in an updated drug glossary based on synonym mining/> Semantically related or not tags/> , if the semantic association exists then , otherwise/> ;Use nonlinear activation function/> Get prediction results based on BERT semantic embedding representation/> , in this embodiment, the sigmoid activation function is used, expressed as , the loss function uses the binary cross-entropy loss function, expressed as ; Use manual annotation to obtain positive training samples, and use random extraction to obtain negative training samples to obtain a training set, conduct BERT model training, and optimize and adjust the BERT semantic embedding representation.

Step S32: Obtain structural embedding representations of external drug terms in the electronic medical record and standard drug term pairs in the drug term database updated based on synonym mining through the graph convolutional neural network model;

Specifically, a candidate association relationship is established between external drug terms in the real-world electronic medical record and drug terms in the drug term database updated based on synonym mining: the TF-IDF value of each word in each drug term is calculated, and then the external drug terms in the electronic medical record are obtained. The vector representation of drug terms and each drug term in the drug term database updated based on synonym mining is used to calculate their previous similarity using cosine similarity, and a similarity threshold is set. If the above similarity is greater than the above similarity threshold, the electronic A candidate association relationship is established between the external drug terms in the medical record and the drug terms in the corresponding drug term database, expressed as

, an example is shown in Figure 4, esomeprazole establishes a candidate association with esomeprazole magnesium, and esomeprazole magnesium establishes a candidate association with esomeprazole magnesium.

Convert external drug terms in electronic medical records into sequences , convert drug terms in the drug term database updated based on synonym mining into sequences /> , use the BERT model trained in step S31 to calculate the semantic embedding representation of the above sequence, and use the semantic embedding representation corresponding to the starting character [CLS] as the initialization node embedding representation of the corresponding drug term;

Embedding the above initialization nodes represents the input graph convolutional neural network model. Specifically, the graph convolutional neural network model includes L layers. In this embodiment, L=10 is taken, where the input of the l-th layer includes two parts. The first part is

dimensional node embedding representation matrix , where n represents the number of nodes, which is the sum of external drug terms in the electronic medical record and drug terms in the drug term database updated based on synonym mining,/> Represents the node embedding representation dimension of the l-th layer, and the second part is/> Dimensional adjacency matrix A, the output of the l-th layer is used as the node embedding representation matrix of the l+1-th layer, which is obtained through the normalized graph Laplace transform. The formula is:

in As a nonlinear activation function, you can use the sigmoid activation function,/> ,I is the identity matrix,/> is a diagonal matrix, and the values of the elements on the diagonal are/> ,/> is the weight matrix of layer l;

The value of the above adjacency matrix A, specifically, if there is a drug term in the drug term database updated based on synonym mining to/> side, then/> The value is 1, otherwise the value is 0; if there is an external drug term from the electronic medical record/> To drug terms in updated drug terminology database based on synonym mining/> side, then/> The value is the value of their similarity in the above candidate association relationship;

The output of the final L-layer graph convolutional neural network model is used as the node embedding representation of drug terms in the drug terminology database updated based on synonym mining and external drug terms in the electronic medical record, from which each drug term database updated based on synonym mining is obtained. standard drug terminology and external drug terms in each electronic medical record/> The node embedding representation of , the product of the node embedding representations of the two is used as the representative/> and/> Structural embedding representation of correlation, denoted as/> ;

The marginal-based distance loss function is used to optimize the graph convolutional neural network model. The loss function formula is:

in Represents standard drug terms in the drug term database updated based on synonym mining/> and external drug terms in each electronic medical record/> The structural embedding represents the distance function /> is a hyperparameter representing the marginal value that distinguishes positive and negative samples,/> represent the positive and negative sample sets respectively; the distance function represented by the structural embedding used in this embodiment is the Euclidean distance, that is,/> , in this embodiment, take/> .

Step S33: Embedding the semantics output in step S31 to represent and the structure embedding representation output by step S32/> Spliced together, expressed as/> , use the above representation as the input of the multi-layer perceptron. The multi-layer perceptron includes multiple fully connected hidden layers and a single-node output layer. The output layer is expressed as/> ;Put the output vector of the multi-layer perceptron through the nonlinear activation function/> Convert to scalar, finally get the external drug terms in each electronic medical record/> Standard drug terms in updated drug terminology library based on synonym mining/> The correlation probability of As the output, the sigmoid activation function is used in this embodiment, expressed as ;The loss function adopts the same binary cross-entropy loss function as the BERT model, expressed as , of which/> External drug terminology for electronic medical records/> Standard drug terms in updated drug terminology library based on synonym mining/> Tags associated or not, if/> and/> If the association exists/> , otherwise/> .

Step S4: Use the association prediction model to predict the association results between external drug terms in the electronic medical record and standard drug terms in the drug terminology database, and establish an association between external drug terms in the real-world electronic medical record and standard drug terms in the drug terminology database.

As shown in Figure 5, the present invention also provides an embodiment of a standardized association system for drug terms in electronic medical records implemented based on the above method. The system includes:

The electronic medical record drug term input module is used to obtain all external drug terms to be standardized in the electronic medical record;

The drug term database synonym mining update module is used to construct a corpus, obtain a drug term list for synonym mining from the corpus; obtain a synonym set for each standard drug term in the drug term database; train a synonym set classifier to obtain all Classification prediction results of each drug term in the drug term list and the synonym set in the drug term database are obtained, and all synonym sets updated based on synonym mining are obtained according to the preset probability threshold, and the drug term database is updated;

The candidate association relationship establishment module is used to establish a candidate association relationship between external drug terms in the electronic medical record and drug terms in the updated drug term database based on similarity calculation;

The semantic embedding representation module is used to combine external drug terms and their pinyin character sequences in electronic medical records, standard drug terms and their pinyin character sequences in the updated drug term database, combined with starting characters and delimiter characters to form an associated drug term pair character sequence. , input the pre-trained language model and obtain the semantic embedding representation;

The structural embedding representation module is used to use the semantic embedding representation of external drug terms in the electronic medical record and the updated drug terminology database as the initialization node embedding representation of the corresponding drug terms, and input the graph convolutional neural network model to obtain the corresponding drug. The node embedding representation of the term is a product of the node embedding representation of the external drug term and the standard drug term as a structural embedding representation;

The association prediction module is used to train an association prediction model based on semantic embedding and structural embedding based on the updated drug term database and external drug terms in the electronic medical record; use the association prediction model to predict and obtain the external drug terms and drug terms in the electronic medical record Association results for standard drug terms in the library.

Corresponding to the embodiments of the method for standardized association of drug terms in electronic medical records, the present invention also provides an embodiment of a device for standardized association of drug terms in electronic medical records. The device for standardizing drug terminology association in electronic medical records provided by an embodiment of the present invention includes a memory and one or more processors. The memory stores executable code. When the processor executes the executable code, it is used to implement The standardized association method of drug terms in electronic medical records in the above embodiment.

An embodiment of the present invention also provides a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the method for standardized association of drug terms in electronic medical records in the above embodiment is implemented.

The computer-readable storage medium may be an internal storage unit of any device with data processing capabilities as described in any of the foregoing embodiments, such as a hard disk or a memory. The computer-readable storage medium can also be an external storage device of any device with data processing capabilities, such as a plug-in hard disk, a smart memory card (SMC), an SD card, or a flash memory card equipped on the device. (Flash Card) etc. Furthermore, the computer-readable storage medium may also include both an internal storage unit and an external storage device of any device with data processing capabilities. The computer-readable storage medium is used to store the computer program and other programs and data required by any device with data processing capabilities, and can also be used to temporarily store data that has been output or is to be output.

The above are only preferred embodiments of one or more embodiments of this specification, and are not intended to limit one or more embodiments of this specification. Within the spirit and principles of one or more embodiments of this specification, Any modifications, equivalent substitutions, improvements, etc. shall be included in the scope of protection of one or more embodiments of this specification.

Claims (2)

1. A method for standardized association of drug terms in electronic medical records, which is characterized by including the following steps: S1. Enter a drug term database to obtain a synonym set for each standard drug term; the drug term database includes a drug term set and a relationship set between drug terms, and the drug term set includes a hypernym or a non-hypernym. Standard drug terminology and synonyms of standard drug terminology; the relationship set includes hyponym relationships and synonym relationships; S2, obtain an updated drug terminology database based on synonym mining, including: Constructing a corpus for synonym mining and obtaining a list of drug terms from said corpus; Train a synonym set classifier to obtain classification prediction results for each drug term in the drug term list and the synonym set in the drug term database, and obtain all synonym sets updated based on synonym mining according to a preset probability threshold; the synonym set During the training process of the classifier, the probability that the drug term to be classified belongs to the synonym set is predicted based on the change of the set unity score. The set unity score calculation method is: calculate the embedding representation of each term in the set, and input the embedding representation into the full connection The neural network model obtains a new term representation, adds all new term representations and takes the average to obtain an initialized term set representation, and inputs the initialized term set representation into the fully connected neural network model to obtain a set unity score; The training set generation method of the synonym set classifier includes: using random extraction to extract a drug term from the synonym set, and combining it with a set of other drug terms in the synonym set to generate a positive training sample; for each positive training sample, match multiple Negative training samples, the negative training samples are obtained by extracting from the drug term database after excluding the drug terms in the synonym set; The drug term database is updated according to all the synonym sets updated based on synonym mining, specifically: if the synonym set of the standard drug term as a hypernym is updated, the corresponding synonyms are established as synonyms with the standard drug term as a hypernym. At the same time, establish a hyponym association between the corresponding synonyms and all the standard drug terms associated with the standard drug term as a hypernym; if the synonym set of the non-hypernym standard drug term is updated, associate the corresponding synonym with the non-hypernym standard drug term. Establish synonym associations for standard drug terms in hypernyms; S3, based on the updated drug term database and external drug terms in the electronic medical record, train an association prediction model based on semantic embedding and structural embedding, including: The semantic embedding representation of the pairs of external drug terms in the electronic medical record and the standard drug terms in the updated drug term database is obtained through the pre-trained language model, specifically: the external drug terms and their pinyin character sequences, the standard drug terms and The pinyin character sequence, combined with the starting character and the separator character, forms a character sequence of associated drug term pairs, which is input into the pre-trained language model to obtain a semantic embedding representation; Adjust the pre-trained language model, specifically: use the semantic embedding representation of the starting character as the independent variable, and the dependent variable is the label of whether the external drug terms in the electronic medical record are semantically related or not in the updated drug term database; use The nonlinear activation function obtains prediction results based on semantic embedding representation; manual annotation is used to obtain positive training samples, and random extraction is used to obtain negative training samples to obtain a training set. Pre-training language model training is performed to optimize and adjust the semantic embedding representation; Obtain the structural embedded representation of the pair of external drug terms in the electronic medical record and the standard drug terminology in the updated drug terminology database through the graph convolutional neural network model, specifically: combining the external drug terminology with the updated drug terminology database Drug terms establish candidate association relationships based on similarity calculation, including: calculating the TF-IDF value of each word in the drug term, obtaining the vector representation of each drug term in the external drug term in the electronic medical record and the updated drug term database, calculating The similarity between the two drug terms. If the similarity is greater than the preset similarity threshold, then a candidate association relationship is established between the external drug term in the electronic medical record and the drug term in the corresponding drug term database; the external drug is The term and the semantic embedding representation of the drug term in the updated drug term database are respectively used as the initialization node embedding representation of the corresponding drug term. The graph convolutional neural network model is input to obtain the node embedding representation of the corresponding drug term. The external drug is The product of the node embedding representation of the term and the standard drug term is used as the structure embedding representation; The input of each layer in the graph convolutional neural network model includes two parts. The first part is the node embedding representation matrix, and the second part is the adjacency matrix. The output of each layer is used as the node embedding representation matrix of the next layer. Through normalization Obtained from the graph Laplace transform of If there is an edge of another drug term, then the corresponding value is 1, otherwise the value is 0; if there is an edge from the external drug term in the electronic medical record to the drug term in the updated drug term database, then the corresponding value is candidate association The similarity value in the relationship; The semantic embedding representation and the structure embedding representation are spliced, and the spliced representation is input into a multi-layer perceptron. The multi-layer perceptron includes multiple fully connected hidden layers and a single-node output layer. The multi-layer perceptron is The output of is converted into a scalar through a nonlinear activation function to obtain the association probability between external drug terms in each electronic medical record and standard drug terms in the updated drug term database; S4: Use the association prediction model to predict the association results between external drug terms in the electronic medical record and standard drug terms in the drug term database. 2. A standardized association system for drug terms in electronic medical records implemented based on the method of claim 1, characterized in that it includes: The electronic medical record drug term input module is used to obtain all external drug terms to be standardized in the electronic medical record; The drug term database synonym mining update module is used to construct a corpus, obtain a drug term list for synonym mining from the corpus; obtain a synonym set for each standard drug term in the drug term database; train a synonym set classifier to obtain all Classification prediction results of each drug term in the drug term list and the synonym set in the drug term database are obtained, and all synonym sets updated based on synonym mining are obtained according to the preset probability threshold, and the drug term database is updated; The candidate association relationship establishment module is used to establish a candidate association relationship between external drug terms in the electronic medical record and drug terms in the updated drug term database based on similarity calculation; The semantic embedding representation module is used to combine external drug terms and their pinyin character sequences in electronic medical records, standard drug terms and their pinyin character sequences in the updated drug term database, combined with starting characters and delimiter characters to form an associated drug term pair character sequence. , input the pre-trained language model and obtain the semantic embedding representation; The structural embedding representation module is used to use the semantic embedding representation of external drug terms in the electronic medical record and the updated drug terminology database as the initialization node embedding representation of the corresponding drug terms, and input the graph convolutional neural network model to obtain the corresponding drug. The node embedding representation of the term is a product of the node embedding representation of the external drug term and the standard drug term as a structural embedding representation; The association prediction module is used to train an association prediction model based on semantic embedding and structural embedding based on the updated drug term database and external drug terms in the electronic medical record; use the association prediction model to predict and obtain the external drug terms and drug terms in the electronic medical record Association results for standard drug terms in the library.