Entity association method for multilingual medical cases

Abstract

The invention belongs to the technical field of computer application, and particularly relates to a multilingual medical case entity association method, which comprises the following steps of: performing head entity and tail entity identification on an original sentence in a medical case; the method comprises the following steps: constructing a dictionary containing entities corresponding to different languages based on medical term contrast data of different languages, translating the entities of different languages into a unified target language by using an online translation API, mapping the entities of different languages to the same embedding space by using a pre-trained multi-language word embedding model, and performing multi-language word embedding. Calculating the similarity of entities of different languages by using a cosine similarity measurement method, and performing entity alignment on the entity with the highest similarity score; constructing a knowledge graph according to the relationship between the entities so as to realize entity association; according to the method, the relation accuracy, practicability and inclusiveness of the knowledge graph are improved, and medical informatization and globalization research and application are promoted.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing technology, and in particular relates to an entity association method for multilingual medical records. Background Art

[0002] With the continuous advancement of globalization, the research and application of multilingual medical data has gradually become an important topic in the field of medical informationization. As an important part of medical data, medical records contain rich diagnosis, treatment and prognosis information, which is of great significance for clinical decision support, medical education and scientific research. Medical records (i.e., medical records, treatment records, etc.) from different countries and regions have accumulated in large quantities in different languages. These multilingual medical records are not only an important resource for medical research, but also a key reference in clinical diagnosis and treatment. However, there are a large number of entities in multilingual medical records, such as diseases, symptoms, examinations, drugs, etc. Due to the language diversity and complexity of medical terminology, how to effectively extract and associate entity information from multilingual medical records has become an urgent problem to be solved.

[0003] The medical records knowledge base is built on the basis of high-quality cases in the field and is a concrete manifestation of clinical knowledge in traditional Chinese medicine. In recent years, the real-world data-driven model has emerged in the clinical decision support system CDSS, and the knowledge base based on the good medical record data in the real world has gradually received attention. Traditional Chinese medicine medical records and medical record data have always occupied an important position in the field of traditional Chinese medicine. Li Xinxia et al. built a knowledge base based on the spleen and stomach medical records of old Chinese medicine practitioners based on ontology. Yu Qi et al. of the China Academy of Chinese Medical Sciences built a clinical ontology based on medical records, realizing the automatic construction of medical record databases, intelligent terminology standardization and analysis and mining functions. At present, a cloud platform of ancient and modern medical records with nearly 300,000 medical records of ancient and modern famous doctors has been built, and a sharing mechanism has been established to encourage everyone to participate in co-construction. The TCM-CDSS clinical decision-making support system for traditional Chinese medicine has been developed, combining Agent technology and intelligent engines to realize the recommendation of similar medical records and prescriptions. Tong Ruan et al. extracted entities, relationships and attributes from semi-structured texts such as encyclopedia websites based on the unified medical language system UMLS to train conditional random field models, thereby automatically extracting knowledge from the electronic medical record text of traditional Chinese medicine to build a knowledge base.

[0004] Early studies mainly used rule-based methods for entity association. This method matches and associates entities in medical records by formulating a series of rules, such as grammatical rules and semantic rules. However, this method has the following disadvantages: the rule formulation is complex and requires researchers with rich professional knowledge to manually write, which is time-consuming and labor-intensive; the rule adaptability is poor and it is difficult to deal with medical records in different languages ​​and different expressions; the scalability is poor, and when new entity types are added to medical records, the rules need to be rewritten. In order to overcome the limitations of rule-based methods, researchers have proposed statistical-based methods. This type of method learns the association probability between entities by training the corpus, thereby achieving entity association. However, statistical-based methods still have the following disadvantages: they rely on a large amount of high-quality annotated data, and it is difficult to obtain annotated data for multilingual medical records; the model has weak generalization ability, and the association effect is poor for new entities or new relationships that have not been seen; the computational complexity is high and it is difficult to meet real-time requirements.

[0005] In recent years, deep learning technology has achieved remarkable results in the field of natural language processing. Researchers have tried to apply deep learning to multilingual medical record entity association tasks, such as using convolutional neural network (CNN) and recurrent neural network (RNN) models for entity association. However, these methods still have the following problems: the model training process relies on a large amount of annotated data, and it is difficult to obtain multilingual medical record annotated data; the model is difficult to capture long-distance dependencies, which affects the entity association effect; the model has poor interpretability, which is not conducive to clinicians' understanding and acceptance.

[0006] In recent years, researchers have proposed a joint model based on graph convolutional networks (GCNs) for extracting Chinese medical entities and their relations. The main purpose of this approach is to solve common entity and relation extraction problems in the medical field, such as nested entities, overlapping relations, etc. Specifically, the model introduced in the article uses the RoBERTa pre-trained language model and a single-module global pointer to handle the task of entity and relation extraction in Chinese medical text. RoBERTa is a powerful pre-trained language model that can understand the complex structure and semantics of the text. The single-module global pointer is used to transform the entity relationship extraction problem into a global pointer linking problem, so as to better capture the relationship between entities. The advantage of this model is that it can integrate global information and improve the accuracy of entity and relation extraction.

[0007] In the prior art, information extraction technology based on natural language processing (NLP) has been widely used in monolingual medical case processing. The core tasks of information extraction include entity extraction (EE) and relationship extraction (RE). Entity extraction is to identify medical entities from text, such as disease names, drugs, symptoms, etc.; relationship extraction is to determine the relationship between these entities, thereby constructing structured knowledge of medical cases. The KECI framework can also be used to jointly extract entities and relationships in biomedical texts. Compared with the general news field, extracting information (IE) from biomedical texts requires more extensive domain knowledge. However, the prior art is mainly limited to processing monolingual texts, and the integrated processing of multilingual medical cases still faces the following significant shortcomings:

[0008] The first is the semantic differences between languages. Existing monolingual models cannot fully handle the semantic differences between different languages, especially in the translation of medical terms and language feature processing, which limits the accuracy of entity extraction and relationship recognition.

[0009] Secondly, the generalization ability of language models is insufficient. Most current NLP models are trained for a specific language and it is difficult to maintain consistent performance in a multilingual context, especially for multilingual medical records with complex structures and large data volumes. The generalization ability of traditional models is relatively weak.

[0010] There is also the language dependency of medical knowledge. Existing information extraction technologies fail to fully consider the cross-language transmission of medical knowledge. Medical terms in different languages ​​may differ significantly in expression and semantics, making it impossible for models trained directly on a single language to accurately capture the relationship between entities when processing multilingual medical records.

[0011] Finally, datasets and annotation rules are rarely used publicly. In recent years, with the development of artificial intelligence technology, especially machine learning and transformer models, NLP has made significant progress in processing clinical free text. However, datasets and annotation rules used for these tasks are rarely used publicly. Most studies fail to provide trained models, datasets, or directly usable tools, and even fewer studies apply these technologies to clinical practice.

[0012] In response to these problems, multilingual entity association methods have gradually become a research hotspot. By integrating natural language processing and data mining technology, this method can not only effectively extract entity information from multilingual texts, but also accurately identify and associate entity relationships in texts of different languages, thereby achieving unified management and in-depth analysis of multilingual medical records. This method can greatly improve the efficiency and accuracy of multilingual medical record processing and promote the development of medical informatization. Summary of the invention

[0013] The present invention aims to overcome the deficiencies in the prior art and provide an entity association method for multilingual medical records.

[0014] In order to achieve the purpose of the present invention, the present invention will be implemented by adopting the technical solution described below.

[0015] A method for associating entities in multilingual medical records comprises the following steps:

[0016] S1. Preprocess the original sentences in the medical records;

[0017] S2. Recognize the head entity and tail entity of the original sentence in the medical record, and generate triples based on the recognized head entity and tail entity; where:

[0018] The head entity recognition method comprises the following steps:

[0019] S211, input the sentence after text preprocessing in step S1 into the pre-trained language model to encode contextual semantic information, and obtain a semantic information feature vector; wherein: the semantic information feature vector is:

[0020] H (0) =BERT(S sw ),

[0021] In the formula, H (0) is the input sentence S sw Contextual semantic information vector at layer 0;

[0022] The encoding is achieved through a multilingual vocabulary table;

[0023] S212, using the semantic information feature vector obtained in step S211 to train two linear classifiers to determine the start position and end position of the head entity of the triple information; wherein: the recognition formula of the head entity is:

[0024]

[0025] In the formula, and Respectively represent the sentence S sw The i-th token in is the probability value of the start position and the end position of the head entity; the probability value is a value between 0 and 1, and the larger the value, the more likely the position is the start position or the end position of the head entity; is the contextual semantic information vector of the i-th token after L layers of BERT processing; W s and W o is a trainable weight matrix; b s and b o is the bias term;

[0026] The method for identifying the tail entity comprises the following steps:

[0027] S221, parsing the original sentences in the medical records, constructing a dependency tree for each sentence in the medical records, and generating an adjacency matrix of the sentence dependency tree;

[0028] S222, according to the sentence dependency tree adjacency matrix generated in step S221, establish a mapping relationship from the original sentence to the subword sequence, generate a mapping table from the subword sequence to the original sentence, and construct a subword dependency tree adjacency matrix that reflects the dependency relationship between subwords;

[0029] S223, input the subword dependency tree adjacency matrix constructed in step S222 into a bidirectional graph convolutional neural network to perform sentence structure information modeling to obtain a feature vector of the structure information; wherein: the formula of the structure information feature vector is:

[0030] H (l+1) =δ(D ％-1 / 2 A ％ D ％-1 / 2 H (l) W (l) ),

[0031] Among them, H (l) is the feature matrix of the lth layer: W (l) is the weight matrix of the 1lth layer; δ is the activation function; A ％ is the normalized adjacency matrix; D ％ is the degree matrix;

[0032] S224, concatenate the semantic information feature vector obtained in step S211 and the structural information feature vector obtained in step S223 to form a joint feature vector, and use the joint feature vector to train two linear classifiers to determine the start position and end position of the tail entity of the triple information; wherein: the recognition formula of the tail entity is:

[0033]

[0034] In the formula, and Respectively represent the sentence S sw The i-th token is the probability value of the start position and the end position of the tail entity; h i is the combined feature representation of the i-th token in the combined feature vector; W p and W q is a trainable weight matrix; b p and b q is the bias term;

[0035] S3. Based on the medical terminology comparison data in different languages, a dictionary containing corresponding entities in different languages ​​is constructed;

[0036] S4. Use the online translation API to translate entities in different languages ​​into a unified target language, and use the pre-trained multilingual word embedding model to map entities in different languages ​​to the same embedding space;

[0037] S5. Calculate the similarity of entities in different languages ​​using the cosine similarity measurement method, and perform entity alignment on the entities with the highest similarity scores;

[0038] S6. Build a knowledge graph based on the relationship between entities to achieve entity association.

[0039] As a preferred embodiment of the present invention, the text preprocessing process comprises the following steps:

[0040] S11. Text cleaning: remove irrelevant characters and symbols from medical records;

[0041] S12, using a word segmentation tool to perform fine segmentation on the text cleaned in step S11;

[0042] S13, using a part-of-speech tagging tool to assign correct part-of-speech tags to different words segmented in step S12.

[0043] As a preferred embodiment of the present invention, the pre-trained language models include: BERT, GPT, Transformer-XL, XLNet, RoBERTa, ALBERT, ELECTRA, DeBERTa, ERNIE and ERNIE2.0 models.

[0044] As a preferred embodiment of the present invention, the parsing is a process of performing natural language processing using the CoreNLP toolkit.

[0045] As a preferred solution of the present invention, the multilingual word embedding model includes Multilingual BERT, LASER, MultiFiT, XLM and XLM-RoBERTa.

[0046] A method for searching for multilingual medical record related entities comprises the following steps:

[0047] Step S51, constructing a data set of medical records in different languages, and extracting entities and relationships between medical records in different languages ​​by using the entity association method of multilingual medical records;

[0048] Step S52: construct a first entity set and a first relationship set between medical records in different languages, establish an initial medical record graph, and construct a second entity set between medical records in different languages ​​using core key entities;

[0049] Step S53: adding the second entity set to the initial medical case graph, performing natural language processing on the entities of medical cases in different languages, and forming a unified label;

[0050] Step S54: Based on the unified label as a subset, the matching degree between different subsets and the set is calculated, and the subset with a matching degree greater than or equal to a preset value is used as the search result set.

[0051] As a preferred embodiment of the present invention, the entities in step S51 include symptoms, treatment plans and acupoints.

[0052] As a preferred embodiment of the present invention, the first entity set FE = {medical case, disease, treatment plan, acupoint}, the first relationship set FR = {suffering from, receiving, using}, and the initial medical case map G = {FE, FR}.

[0053] As a preferred solution of the present invention, the method for constructing the second entity set is to extract the core key entities in medical records in different languages: acupoints, and construct a second entity set of patients and symptoms in medical records, recorded as the second entity set SE = {medical records, acupoint locations}.

[0054] Beneficial Effects

[0055] 1. Enrich the content of the multilingual medical case knowledge graph

[0056] Expanding the types and quantity of entities: This method can more accurately extract entities from long texts of multilingual medical records, such as the four diagnostic methods of traditional Chinese medicine, traditional Chinese medicine dialectics, acupuncture treatment, etc., thereby greatly enriching the types and quantity of entities in the knowledge graph of multilingual medical records. Multimodal technology can provide a richer representation of entities and relationships in the knowledge graph by integrating multiple types of data such as text, images, and sounds, and enhance the expression ability of knowledge. For example, in the human body structure diagram, locate Tianshu "Tianshu, one inch and five fen away from the gate of the uterus, two inches on both sides of the umbilicus, and the sunken middle", and attach descriptions in different languages ​​such as Chinese and English.

[0057] 2. Improving the accuracy of relationships in the knowledge graph

[0058] Enhance the accuracy of relationship extraction: By jointly extracting entity relationships, this method can more accurately identify complex relationships between entities in multilingual medical records, such as the correspondence between diseases and symptoms, the correspondence between acupoints and treatment effects, etc., thereby improving the accuracy of relationship descriptions in multilingual medical record knowledge graphs. By using complementary information between multimodal data, the ambiguity problem in single-modal data can be resolved, and the accuracy of entity recognition and relationship extraction can be improved. For example, in the process of studying the mechanism of acupuncture in lowering blood sugar, not only can the mechanism of action of commonly used acupoints in China, such as spleen, diaphragm, and Zusanli, be retrieved, but also the latest international research on the principle of acupuncture in lowering blood sugar can be extracted, thus enhancing the accuracy of entity relationships in multilingual medical records.

[0059] 3. Improving the practicality of knowledge graphs

[0060] Enhance the query and reasoning capabilities of the graph: With the continuous enrichment and improvement of entities and relationships in the multilingual medical case knowledge graph, users can obtain the required medical cases through more accurate query and reasoning. For example, doctors can quickly find corresponding multilingual medical cases and treatment methods in the graph according to the patient's symptoms, improving diagnosis and treatment efficiency and multilingual medical capabilities. The fusion of multimodal information can provide more contextual information for the reasoning of the knowledge graph, enhance the reasoning and prediction capabilities of the knowledge graph, and assist in the automatic update and maintenance of the knowledge graph. For example, the latest domestic and foreign cutting-edge research can be extracted to maintain the timeliness and advancement of the knowledge graph.

[0061] 4. Improving the inclusiveness of medical services

[0062] Improve the multilingual medical capabilities of the graph: By using the proxy model of the large model, covering patients and medical case graphs in different languages, significantly improve the user's multilingual medical capabilities, and provide rich data support for medical services worldwide. This entity association method of multilingual medical cases not only increases the feature dimension and information richness, but also helps to improve the efficiency and inclusiveness of medical services. Doctors can refer to domestic and foreign medical cases and cutting-edge research to provide more personalized and advanced treatment plans for different races and different diseases.

[0063] 5. Promote research and application of medical informationization and globalization

[0064] Promoting the innovative development of medical informationization and globalization: This method provides powerful data support and analysis tools for the research of medical informationization. Through in-depth mining of long texts of multilingual medical records, multilingual medical record information, diagnosis and treatment plans, etc. can be discovered, providing strong support for the innovative development of medical informationization. A large amount of literature has been accumulated in the field of medical informationization. Multimodal extraction technology can automatically identify entities and relationships related to multilingual medical records from the literature, construct a knowledge graph, and reveal the internal connections and development trends of multilingual medical records. At the same time, the construction and application of multilingual medical record knowledge graphs will also help promote the development and research of intelligent and internationalized traditional Chinese medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 The present invention is a flowchart of a method for associating entities in multilingual medical records. DETAILED DESCRIPTION

[0066] The present invention is further described in conjunction with the accompanying drawings and embodiments.

[0067] As an embodiment of the present invention, Figure 1As shown, a method for entity association of multilingual medical records includes the following steps:

[0068] 1. Text Preprocessing

[0069] Taking the medical records in "16 Japanese Acupuncture and Moxibustion Books" as an example, the preprocessing step is crucial because the long text of the medical records contains professional terms, ancient Chinese expressions and complex sentences. The technical solution includes text cleaning (removing irrelevant characters and symbols), word segmentation (using word segmentation tools for fine segmentation), and part-of-speech tagging (assigning correct part-of-speech tags to different words). For example, in "Da Ming Zhuo Zhou Axis of Acupuncture Methods", "coughing continuously, the lung point, Xu Ying Tian Tu acupoint; red and astringent urine, Dui Duan Du Tai Yang meridian." The Chinese word segmentation tool Jieba is used for word segmentation. The database contains words such as "cough", "urination", "lung point", "Dui Duan", etc. The word segmenter will segment according to the word database. The word segmentation result is: cough|continuous|lung point|Xu Ying|Tian Tu acupoint|urination|red and astringent|Dui Duan|Du Tai Yang meridian. After word segmentation, the part-of-speech tagging tool NLPIR is used to tag the word segmentation results, that is, to add a part-of-speech tag to each word, such as noun (N), verb (V), adjective (A), etc. The results of word segmentation and part-of-speech tagging are: cough (v) | continuous sound (v) | Feishu (nz) | Xuying (n) | Tiantu point (nz) | urination (n) | red and astringent (n) | Duiduan (nz) | Du (a) | Taiyangjing (n). Here, "nz" means proper noun, such as name of a person, place name, acupuncture point name, etc.

[0070] Different word segmentation methods are used for medical records in different languages. For English medical records, special characters and extra spaces in the text are first removed, and then the word segmentation tool NLTK is used for word segmentation. Finally, the word segmentation tool NLTK is used to tag the word segmentation results. For example, in "Acupuncture Collection": "The patient has delayed menstruation, yellow complexion, and vomiting during pregnancy: the points are Quchi (LI11), Zhiqiao (GB37), Zusanli (ST36), and Sanyinjiao (SP6).". The word segmentation result is: "the|patient|has|delayed|menstruation|yellow|complexion|and|vomiting|during|pregnancy|the|points|are|Quchi|LI11|Zhiqiao|GB37|Zusanli|ST36|and|Sanyinjiao|SP6".

[0071] Through the above steps, medical records in different languages ​​can be effectively preprocessed, which is crucial for subsequent steps such as named entity recognition and relationship extraction.

[0072] 2. Entity Alignment

[0073] The semantic information feature vector of each sentence is obtained by pre-training the language model. The semantic information feature vector is used to train two linear classifiers to determine the start and end positions of the head entity. For a sentence S = {w1, w2, ..., w n}, where w t is the tth word in sentence S. The sentence is segmented using the Jieba algorithm to obtain the segmentation sequence S sw .

[0074] The BERT pre-trained model is used to obtain the initial representation of entities, attributes, and categories, and fully obtain semantic information. Semantic information feature vector:

[0075] H (0) =BERT(S sw ),

[0076] Among them, H (0) It is the contextual semantic information vector of the input sentence at layer 0. The semantic information feature vector of each sentence is obtained by pre-training the language model, and two linear classifiers are trained using the semantic information feature vector to determine the start and end positions of the head entity.

[0077] Head entity recognition formula:

[0078]

[0079] in, and Respectively represent the sentence S sw The i-th token in H is the probability value of the start position and end position of the head entity; it is a value between 0 and 1, and the larger the value, the more likely the position is to be the start position or end position of the head entity. i (L) is the contextual semantic information vector of the i-th token after L layers of BERT processing; W s and W o is a trainable weight matrix; b s and b o is the bias term.

[0080] The CoreNLP toolkit is used to perform natural language processing on the medical case text, build a dependency tree for each sentence, and generate a dependency tree adjacency matrix T. The elements in the matrix T ij 1 represents word wi With the word w j There is a dependency relationship between them, and 0 means there is no dependency relationship. A mapping relationship is established from the original sentence to the subword sequence, and a mapping table M from the subword sequence to the original sentence is generated. A new adjacency matrix T is constructed. sw To reflect the dependency relationship between subwords.

[0081] The reconstructed dependency tree adjacency matrix is ​​processed, and the bidirectional graph convolutional neural network Bi-GCN is used to model the structural information to obtain the structural information feature vector.

[0082] Structural information feature vector formula:

[0083] H (l+1) =δ(D ％-1 / 2 A ％ D ％-1 / 2 H (l) W (l) )

[0084] Among them, H (l) is the feature matrix of the lth layer: W (l) is the weight matrix of the 1lth layer; δ is the activation function; A ％ is the normalized adjacency matrix; D ％ is the degree matrix.

[0085] The feature vector reflecting the sentence structure information is obtained, which can be further used to extract triple information. The semantic information feature vector and the structural information feature vector are combined to more accurately identify the head entity and tail entity in the sentence, thereby solving the triple overlap problem and processing long and difficult sentences.

[0086] The semantic information feature vector is concatenated with the structural information feature vector to form a joint feature vector, and the joint feature vector is used to train two linear classifiers to determine the start and end positions of the tail entity. Tail entity recognition formula:

[0087]

[0088] in, and Respectively represent the sentence S sw The i-th token is the probability value of the start position and the end position of the tail entity; h i is the combined feature representation of the i-th token in the joint feature vector; W p and W q is a trainable weight matrix; b p and b q is the bias term.

[0089] For example, the result of word segmentation for a medical case in "16 Japanese Acupuncture and Moxibustion Books" is: the patient|due to|cold|cold|had|headache|fever|and other|symptoms|treated|using|acupuncture|selected|Hegu point|and|Zusanli point|after treatment|symptoms|significantly|relieved|. By training the head entity linear classifier, the start and end positions of the head entity are identified: head entity: "patient", start position: 1, end position: 1. Then the tail entity linear classifier is trained using the joint feature vector to identify the start and end positions of the tail entity: symptoms: "headache", "fever", start positions: 6, 8, end positions: 7, 9; treatment acupoints: "Hegu point", "Zusanli point", start positions: 17, 19, end positions: 18, 20. According to the identified head entity and tail entity, triples are generated: "patient-suffering-headache", "patient-suffering-fever", "patient-treatment-Hegu point", "patient-treatment-Zusanli point".

[0090] The accuracy of entity alignment can be further improved by building rich bilingual dictionaries and training large-scale corpora. Collect existing bilingual comparison data, such as bilingual medical terminology glossaries. Build a dictionary containing corresponding entities in different languages, and use the online translation API to translate entities in different languages ​​into a unified target language. Use a pre-trained multilingual word embedding model (Multilingual BERT) to map entities in different languages ​​to the same embedding space. Use the cosine similarity metric to calculate the similarity of entities in different languages, and align them according to the entities with the highest similarity scores. Finally, manually verify the results of the automatic alignment and correct errors. Identify and annotate the same entities in medical records in different language versions to ensure their uniqueness in the knowledge graph. For example, "patient", "patient", and "patient" should all be aligned as the same entity. Synonyms representing the same concept in different language versions should also be identified and merged in the knowledge graph.

[0091] According to the above triples, we can identify entity sets and relationship sets: entity set: {patient, headache, fever, Hegu point, Zusanli point}, relationship set: {suffering from, treatment}. The above identified entities and their relationships are constructed into a knowledge graph. The initial knowledge graph consists of entity set E and relationship set R, expressed as G = (E, R). The knowledge graph formed by the above example: patient--[suffering from]->headache, patient--[suffering from]->fever, patient--[treatment]->Hegu point, patient--[treatment]->Zusanli point.

[0092] As an embodiment of the present invention, the present invention aims to associate medical records in different languages, wherein existing NLP technology, data and model collaborative optimization technology, etc. are used to digitally manage medical records in different languages, and intelligent medical record search is realized. By extracting the core entities between medical records in different languages, using them as the associated entities of the two medical records, and performing natural language processing on the entities in the intersecting set, the two medical records are connected to improve the processing and analysis efficiency of medical records in different languages. The main method is as follows:

[0093] Step S1: Establish a medical case database in different languages, and extract the relationship and associated entities between two medical cases by using the entity association method of multilingual medical cases;

[0094] For example, a Chinese medical case database and an English medical case database are constructed, and the Chinese medical case database and the English medical case database are used as a general medical case database. The entities extracted from the database by the entity association method of a multilingual medical case include: symptoms, treatment plans and acupoints.

[0095] Step S2: constructing a first entity set and a first relationship set between different languages ​​based on the extracted entities, establishing an initial medical case map, extracting key core related entities, and constructing a second entity set between medical cases in two languages;

[0096] For example: the first entity set FE = {medical case, symptom, treatment plan, acupoint}, the first relationship set FE = {suffering from, receiving, using}, thereby constructing the initial medical case map G = {FE, FR};

[0097] Furthermore, the key core associated entities in the medical records in the two languages ​​are extracted: acupoints, and a second entity set of patients and symptoms in the medical records is constructed, which is recorded as the second entity set SE = {medical records, acupoint locations}.

[0098] Step S3: adding the second entity set and the second relationship set into the initial medical case graph, and subjecting the medical case entities to natural language processing into unified labels;

[0099] Step S4: Based on the unified label as a subset, the matching degree between different subsets and the set is calculated, and the subset with a matching degree greater than or equal to a preset value is used as the search result set.

[0100] For example, the following English paragraph explains the mechanism of acupuncture Tianshu in lowering blood sugar:

[0101] Electroacupuncture(EA)is considered to have potential antidiabeticeffects;however,the role of the pancreatic intrinsic nervoussystem(PINS)inEA-induced amelioration of type 2 diabetes(T2DM)remains unclear.Therefore,weinvestigated whether EA at ST25 exertsany beneficial effects on insulinresistance(IR),inflammation severity,and pancreaticβcell function via thePINS in a rat model of a high-fat diet-streptozotocin(HFD / STZ)-induceddiabetes.To thisend,Sprague Dawley rats were fed with HFD to induce IR,followed by STZ(35 mg / kg,i.p.)injection to establish the T2DMmodel.Afterhyperglycemia was confirmed as fasting glucose level>16.7 mmol / L,the rats were treated with EA(2 mA,2 / 15 Hz)for the next 28 days.Model ratsshowed increased serum glucose,insulin,IR,and TNF-αlevels with a concomitantdecrease inβcell function.Microscopy examination of the pancreas revealedpathological changes in islets,which reverted to near-normal levels after EAat ST25.EA improved islet cell morphology by increasing islet area andreducing vacuolation.EA at ST25 decreased transient receptor potentialvanilloid 1(TRPV1)and increased substance P(SP)and calcitonin gene-relatedpeptide(CGRP)expression.Subsequently,insulin secretion decreased and impairedpancreatic endocrine function was restored through the TRPV1 channel(SP / CGRP)-insulin circuit.EA increased choline acetyltransferase and neuropeptideY expression and controlled inflammation.It also enhanced the cocaine andamphetamine-regulated transcript prepropeptide expression and promotedglucagon-like peptide-1 secretion.Additionally,the electrophysiologicalactivity of PINS during acupuncture(2.71±1.72Hz)was significantly increasedcompared to the pre-acupuncture frequency(0.32±0.37Hz,P<0.05).Thus,our studydemonstrated the beneficial effect of EA onβcell dysfunction via the PINS inrat models of HFD-STZ-induced T2DM.

[0102] In this English medical record, we can extract the keywords related to acupuncture and diseases: electroacupuncture (EA), type 2 diabetes (T2DM), Tianshu (ST25), transient receptor potential vanilloid subtype 1 (TRPV1), (SP) and calcitonin gene-related peptide (CGRP). Set the medical record as a, FE = {a, type 2 diabetes (T2DM), electroacupuncture (EA), Tianshu (ST25)}, FR = {suffering from type 2 diabetes (T2DM), receiving electroacupuncture treatment, using Tianshu acupoint for treatment}, SE = {a, Tianshu}. Based on the first entity set and the second entity set, the keyword entities in this medical record are processed as electroacupuncture (EA), type 2 diabetes (T2DM), Tianshu (ST25), transient receptor potential vanilloid subtype 1 (TRPV1), (SP) and calcitonin gene-related peptide (CGRP), calculate the matching degree of each keyword with the keywords in "Acupuncture and Moxibustion Jia Yi Jing", and extract the subset greater than or equal to the preset value. Through comparison, the record about Tianshu was mentioned in "Acupuncture and Moxibustion Jia Yi Jing", which is "Tianshu, one inch and five fen away from the liver gate, in the middle of the depression two inches on both sides of the navel". That is, Tianshu was included in the final search result set and participated in the search discussion.

[0103] By inputting the search subset to obtain relevant information in the total medical case database, the relevance of the core entities can be increased to improve the efficiency of processing and analyzing medical cases in different languages.

Claims

1. A method for associating entities in medical records in different languages, characterized in that: The steps include: S1. Preprocess the original sentences in medical records in different languages; S2. Recognize the head entity and tail entity of the triple information of the original sentences in the medical records of different languages, and extract the entities of different languages ​​based on the recognized head entity and tail entity of the triple information; wherein: The identification process of the head entity includes the following steps: S211, input the sentence after text preprocessing in step S1 into the pre-trained language model to encode contextual semantic information, and obtain a semantic information feature vector; wherein: the semantic information feature vector is: H (0) =BERT(S sw ), In the formula, H (0) The input sentence S sw Contextual semantic information vector at layer 0; The coding is achieved through medical reference lists between different languages; S212, using the semantic information feature vector obtained in step S211 to train two linear classifiers to determine the start position and end position of the head entity of the triple information; wherein: the recognition formula of the head entity is: In the formula, and Represents the sentence S sw The i-th token in H is the probability value of the start position and the end position of the head entity; the probability value is a value between 0 and 1, and the larger the value, the more likely the position is the start position or the end position of the head entity; H i (L) is the contextual semantic information vector of the i-th token after L layers of BERT processing; W s and W o is a trainable weight matrix; b s and b o is the bias term; The identification process of the tail entity includes the following steps: S221, parsing the original sentences in the medical records, constructing a dependency tree for each sentence in the medical records, and generating an adjacency matrix of the sentence dependency tree; S222, according to the sentence dependency tree adjacency matrix generated in step S221, establish a mapping relationship from the original sentence to the subword sequence, generate a mapping table from the subword sequence to the original sentence, and construct a subword dependency tree adjacency matrix that reflects the dependency relationship between subwords; S223, input the subword dependency tree adjacency matrix constructed in step S222 into a bidirectional graph convolutional neural network to perform sentence structure information modeling to obtain a feature vector of the structure information; wherein: the formula of the structure information feature vector is: H (1+1) =δ(D ％-1 / 2 A ％ D ％-1 / 2 H (1) W (1) ), Among them, H (l) is the feature matrix of the ,th layer: W (l) is the weight matrix of the 1lth layer; δ is the activation function; A ％ is the normalized adjacency matrix; D ％ is the degree matrix; S224, concatenate the semantic information feature vector obtained in step S211 and the structural information feature vector obtained in step S223 to form a joint feature vector, and use the joint feature vector to train two linear classifiers to determine the start position and end position of the tail entity of the triple information; wherein: the recognition formula of the tail entity is: In the formula, and Represents the sentence S sw The i-th token is the probability value of the start position and the end position of the tail entity; h i is the combined feature representation of the i-th token in the combined feature vector; W p and W q is a trainable weight matrix; b p and b q is the bias term; S3. Construct a dictionary containing corresponding entities in different languages ​​based on the medical terminology comparison data in different languages; S4, using the online translation API to translate entities in different languages ​​into a unified target language, and using the pre-trained multilingual word embedding model to map the entities in different languages ​​extracted in step S2 to the same embedding space; S5. Calculate the similarity of entities in different languages ​​using the cosine similarity measurement method, and perform entity alignment on the entities with the highest similarity scores; S6. Build a knowledge graph based on the relationship between entities to achieve entity association.

2. The entity association method for multilingual medical records according to claim 1, characterized in that: The text preprocessing process includes the following steps: S11. Text cleaning: remove irrelevant characters and symbols from medical records; S12, using a word segmentation tool to perform fine segmentation on the text cleaned in step S11; S13, using a part-of-speech tagging tool to assign correct part-of-speech tags to different words segmented in step S12.

3. The entity association method for multilingual medical records according to claim 1, characterized in that: The pre-trained language models include: BERT, GPT, Transformer-XL, XLNet, RoBERTa, ALBERT, ELECTRA, DeBERTa, ERNIE and ERNIE2.0 models.

4. The entity association method for multilingual medical records according to claim 1, characterized in that: The parsing is a process of performing natural language processing using the CoreNLP toolkit.

5. The entity association method for multilingual medical records according to claim 1, characterized in that: The multilingual word embedding models include: MultilingualBERT, LASER, MultiFiT, XLM and XLM-RoBERTa.

6. A method for searching for multilingual medical record related entities, characterized in that: The steps include: Step S61, establishing a medical case database in different languages, constructing a data set of medical cases in different languages, and extracting entities and relationships between medical cases in different languages ​​by using the entity association method of multilingual medical cases as described in claim 1; Step S62, constructing a first entity set FE and a first relationship set FR between medical records in different languages, establishing an initial medical record graph G, extracting key core related entities, and constructing a second entity set SE between medical records in different languages; Step S63: adding the first entity set and the second entity set into the initial medical case graph, performing natural language processing on the entities of the medical case to form a unified label; Step S64: Based on the unified label as a subset, the matching degree between different subsets and the set is calculated, and the subset with a matching degree greater than or equal to a preset value is used as the search result set.

7. A method for searching multilingual medical record related entities according to claim 6, characterized in that: The entities in step S51 include: symptoms, treatment plans and acupoints.

8. The method for searching for multilingual medical record-related entities according to claim 6, characterized in that: The first entity set FE = {medical case, disease, treatment plan, acupoint}, the first relationship set FR = {suffering from, receiving, using}, and the initial medical case map G = {FE, FR}.

9. A method for searching multilingual medical record related entities according to claim 8, characterized in that: The method for constructing the second entity set is to extract the core key entities in medical records in different languages: acupoints, and construct a second entity set of patients and symptoms in medical records, which is recorded as the second entity set SE = {medical records, acupoint locations}.

Images