Method and device for distinguishing the positive and negative nature of disease entities and symptom entities in medical record text

Through the entity recognition and time segmentation model combined with the pre-trained language model, the difficulty of identifying and judging disease and symptom information in medical record text in the prior art is solved, and accurate yin and yang judgment and accuracy improvement are achieved.

CN114218954BActive Publication Date: 2024-11-08BEIJING UNISOUND INFORMATION TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111561244.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-16
Publication Date
2024-11-08
Estimated Expiration
2041-12-16

AI Technical Summary

Technical Problem

The existing Yin-Yang judgment methods rely on templates and dictionaries, resulting in cumbersome maintenance and inability to handle unpredefined situations, and the inability to accurately identify disease and symptom information in medical record text.

Method used

A two-stage method is adopted, firstly, the entity identification model and time segmentation model are used to obtain the entity and the context information of the disease and symptom entities and context information, and then the pre-trained language model is used for fine adjustment to construct training data to judge the yin and yang of the disease and symptoms.

Benefits of technology

Accurate identification and yin-yang judgment of disease and symptom information in medical record text, avoid template and dictionary scale limitations, can handle undefined situations, and continuously improve judgment accuracy through model updates.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114218954B_ABST
    Figure CN114218954B_ABST
Patent Text Reader

Abstract

The present invention provides a method and device for judging the masculine and feminine nature of diseases and symptoms in medical record texts, the judgment method comprising: obtaining disease entities and symptom entities and corresponding context information in target medical record texts; judging the masculine and feminine nature of disease entities and symptom entities in target medical record texts based on the disease entities and symptom entities and corresponding context information. The technical solution provided by the present invention can avoid the problem that the template and dictionary method is limited by the scale of templates and dictionaries and cannot handle undefined situations, and can also continuously improve the accuracy of judgment by increasing corpus and updating the models in the two stages of training.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] One or more embodiments of the present invention relate to the field of medical diagnosis technology, and in particular, to a method and device for determining the positive and negative nature of disease entities and symptom entities in medical record texts. Background Art

[0002] This section is intended to provide background or context for the embodiments of the invention set forth in the claims. The descriptions herein may include concepts that may be explored, but not necessarily concepts that have been previously thought of or explored. Therefore, unless otherwise noted herein, the content described in this section is not prior art with respect to the specification and claims of this application, and is not admitted to be prior art by inclusion in this section.

[0003] Medical records are the sum of data generated by medical personnel during the process of medical activities such as examination, diagnosis, and treatment of the occurrence, development, and outcome of patients' diseases. With the advent of the intelligent era, the diagnosis and treatment of diseases has gradually entered the intelligent era, that is, using existing technical information, as well as the patient's actual condition and related data, to assist in the prediction or pre-diagnosis of the patient's disease, thereby providing certain help to medical staff.

[0004] The existing yin-yang judgment method first needs to construct text templates for different yin-yang situations, and then judge the text fragments containing disease and symptom descriptions through keyword matching and template matching, and use the template matching results to judge the yin-yang of the current disease or symptom description.

[0005] Currently, the vocabulary required to identify diseases or symptoms in texts requires professionals with certain experience to search and organize, build a large dictionary, and regularly maintain and update it, which is time-consuming and labor-intensive. At the same time, the text templates for judging yin and yang also require the same or even larger scale of organization and maintenance, and when expressions outside the template appear, it is impossible to make correct judgments, and the output results are completely dependent on the dictionary and template resources. Summary of the invention

[0006] One or more embodiments of the present specification describe a method, device, and system for determining the positivity of disease entities and symptom entities in a medical record text, which can accurately identify disease and symptom information in a target medical text and make positivity judgments that are consistent with the context.

[0007] The technical solutions provided by one or more embodiments of this specification are as follows:

[0008] In a first aspect, the present invention provides a method for determining the positive and negative nature of disease entities and symptom entities in a medical record text, the method comprising:

[0009] Obtain disease entities and symptom entities and corresponding context information in the target medical record text;

[0010] Based on the disease entity and symptom entity and the corresponding context information, the masculine and feminine nature of the disease entity and the symptom entity in the target medical record text is determined.

[0011] In one example, the step of obtaining the disease entity and symptom entity and the corresponding context information in the target medical record text includes the following steps:

[0012] Obtain an entity recognition model;

[0013] Obtain the time segmentation model;

[0014] The entity recognition model and the time segmentation model are used to obtain disease entities and symptom entities in the target medical record text and corresponding context information.

[0015] In one example, obtaining the entity recognition model includes the following steps:

[0016] Collect medical record texts, and annotate disease entities and symptom entities in the medical record texts;

[0017] Convert the annotated medical record text data into BIO annotation format;

[0018] Initialize the input vector with the word vectors obtained by pre-training the medical record text data in BIO annotation format;

[0019] According to the initialized input vector, a bidirectional recurrent neural network training based on LSTM unit is used to obtain an entity recognition model.

[0020] In one example, obtaining the time period segmentation model includes the following steps:

[0021] Divide the time periods in the medical record text;

[0022] Convert the processed medical record text sample data into BIO annotation format;

[0023] Initialize the input vector with the word vectors obtained by pre-training the medical record text data in BIO annotation format;

[0024] According to the initialized input vector, a bidirectional recurrent neural network training based on LSTM unit is used to obtain a time segmentation model.

[0025] In one example, the method of using the entity recognition model and the time segmentation model to obtain disease entities and symptom entities in the target medical record text and corresponding context information includes the following steps:

[0026] Use the entity recognition model to predict the target medical record text to obtain disease entities and symptom entities;

[0027] The target medical record text is segmented using a time segmentation model, and the beginning part and the middle part of the predicted continuous time segment are regarded as the text content in one time segment; the target medical record text is composed of texts of different time segments;

[0028] Searching the target medical record text for the disease entity and the symptom entity; wherein the disease entity and the symptom entity exist in the corresponding text respectively;

[0029] The disease entity and its corresponding text and the symptom entity and its corresponding text are taken as text pairs, namely, the disease entity and its corresponding context information and the symptom entity and its corresponding context information.

[0030] In one example, judging the masculine and feminine nature of disease entities and symptom entities in a target medical record text based on the disease entities and symptom entities and corresponding context information includes the following steps:

[0031] The disease entity and its corresponding context information and the symptom entity and its corresponding context information are labeled to obtain the positive and negative information of the disease entity and the symptom entity, and to construct training data for judging the positive and negative of the disease entity and the symptom entity; the labeled labels include positive, negative, and cannot be judged;

[0032] Use medical record text to obtain a pre-trained language model;

[0033] The pre-trained language model is used to fine-tune the constructed training data to obtain the positive and negative results of disease entities and symptom entities in the target medical record text.

[0034] In one example, using the pre-trained language model to fine-tune the constructed training data to obtain the positive and negative results of the disease entity and the symptom entity in the target medical record text includes the following steps:

[0035] The labeled training data is concatenated from w and p to form the input U = [u1,…,u n ] and add the label [CLS] before the sequence input from the source end; where w and p represent entity and context information respectively;

[0036] Define Trm() as BERT bio The encoding calculation unit in the code, each word passes through the encoding unit to obtain the encoding representation, which is calculated by the following formula:

[0037]

[0038] BERT bioTake the hidden layer vector representation of the first [CLS] of the model After a layer of fully connected linear transformation:

[0039]

[0040] The probability distribution of the target medical text is given by softmax:

[0041] Prob = softmax(C)

[0042] The maximum probability is taken as the positive or negative identification result of the entity.

[0043] In a second aspect, the present invention provides a device for determining the positive and negative nature of a disease entity and a symptom entity in a medical record text, the device comprising:

[0044] An acquisition module is used to obtain disease entities and symptom entities and corresponding context information in the target medical record text;

[0045] The judgment module is used to judge the masculine and feminine nature of the disease entity and the symptom entity in the target medical record text based on the disease entity and the symptom entity and the corresponding context information.

[0046] In one example, the acquisition module includes:

[0047] A first acquisition unit, used to obtain an entity recognition model;

[0048] A second acquisition unit is used to obtain a time segmentation model;

[0049] The third acquisition unit is used to obtain the disease entity and symptom entity and corresponding context information in the target medical record text by using the entity recognition model and the time period segmentation model.

[0050] In a third aspect, the present invention provides a system for determining the positive and negative nature of disease entities and symptom entities in medical record texts, the system comprising at least one processor and a memory;

[0051] The memory is used to store one or more program instructions;

[0052] The processor is used to run one or more program instructions to execute one or more methods as described in the first aspect.

[0053] In a fourth aspect, the present invention provides a chip, which is coupled to a memory in a system, so that the chip calls program instructions stored in the memory during operation to implement one or more of the methods described in the first aspect.

[0054] In a fifth aspect, the present invention provides a computer-readable storage medium, which includes one or more program instructions, and the one or more program instructions can be executed by the system described in the third aspect to implement one or more methods described in the first aspect.

[0055] The technical solution provided by the embodiment of the present invention can avoid the problem that the template and dictionary method is limited by the scale of templates and dictionaries and cannot handle undefined situations. At the same time, it can also continuously improve the accuracy of judgment by increasing corpus and updating the models in the two stages of training. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 A flow chart of a method for determining the positive and negative nature of disease entities and symptom entities in a medical record text provided by an embodiment of the present invention;

[0057] Figure 2 A flowchart for obtaining disease entities and symptom entities and corresponding context information in the target medical record text;

[0058] Figure 3 A flowchart for obtaining an entity recognition model;

[0059] Figure 4 A flowchart for obtaining a time segmentation model;

[0060] Figure 5 A flowchart for obtaining disease entities and symptom entities and corresponding context information in the target medical record text using the entity recognition model and the time segmentation model;

[0061] Figure 6 A flowchart for determining the masculine and feminine nature of disease entities and symptom entities in a target medical record text based on disease entities and symptom entities and corresponding context information;

[0062] Figure 7 A schematic diagram of the structure of a device for judging the positive and negative nature of disease entities and symptom entities in a medical record text provided by an embodiment of the present invention;

[0063] Figure 8 To obtain the structural diagram of the module;

[0064] Fig. 9 A schematic diagram of the structure of a system for judging the positivity of disease entities and symptom entities in a medical record text provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The present application will be further described in detail below in conjunction with the accompanying drawings and embodiments. It is to be understood that the specific embodiments described herein are only used to explain the relevant invention, rather than to limit the invention. It should also be noted that, for ease of description, only the parts related to the relevant invention are shown in the accompanying drawings.

[0066] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0067] The present invention provides a two-stage method for respectively realizing the recognition of diseases, symptoms and their contexts in medical texts and the classification of their positive and negative nature. In the first stage, a sequence labeling model is used to obtain entity information and context information in the medical texts respectively; in the second stage, a large-scale pre-trained language model trained on medical texts is used to classify the texts constructed by disease entities and symptom entities obtained in the first stage and their corresponding contexts, thereby determining the positive and negative nature of the disease or diagnosis.

[0068] Two points need to be explained here:

[0069] (1) In this application, the medical texts that need to be judged for the positive and negative nature of disease entities and symptom entities are referred to as target medical texts or medical texts to be judged. Medical texts mentioned elsewhere, such as medical texts used for model training, are directly referred to as medical texts.

[0070] (2) The sequence labeling model in the present invention includes an entity recognition model and a time segmentation model, which are used to obtain disease entities and symptom entities in medical texts and their context information, respectively, which will be introduced in the following description:

[0071] Figure 1 The flowchart of the method for judging the yin and yang of disease entities and symptom entities in medical record text is shown as follows: Figure 1 The determination method includes step S10 and step S20:

[0072] Step S10, obtaining disease entities and symptom entities and corresponding context information in the target medical record text.

[0073] In one example, if Figure 2 As shown, the process is mainly implemented through the following three steps S101-S103:

[0074] Step S101, obtaining an entity recognition model.

[0075] In an example, the specific process of obtaining the entity recognition model is as follows: Figure 3 As shown, including the following 4

[0076] Steps S1011-S1014:

[0077] S1011, collect medical record texts, and annotate disease entities and symptom entities in the medical record texts.

[0078] S1012, converting the annotated medical record text data into a BIO annotation format.

[0079] It should be noted here that BIO is a labeling method used for sequence labeling problems, generally called BIO labeling or BIO format labeling. B represents the character at the beginning of the label, I represents other characters in the label except the beginning, and O represents non-label characters. It can be replaced by Chinese characters such as start, non-start, and others.

[0080] S1013, initializing the input vector with the word vectors pre-trained from the medical record text data in BIO annotation format.

[0081] S1014, according to the initialized input vector, use a bidirectional recurrent neural network training based on an LSTM unit to obtain an entity recognition model.

[0082] Step S102: obtaining a time segmentation model.

[0083] In an example, the specific process of obtaining the time segmentation model is as follows: Figure 4 As shown, the following four steps S1021-S1024 are included:

[0084] S1021, dividing the time periods in the medical record text.

[0085] S1022, converting the processed medical record text sample data into a BIO annotation format.

[0086] S1023, initializing the input vector with the word vectors pre-trained from the medical record text data in BIO annotation format.

[0087] S1024: According to the initialized input vector, a bidirectional recurrent neural network training based on an LSTM unit is used to obtain a time segmentation model.

[0088] Step S103, using the entity recognition model and the time period segmentation model, obtain the disease entity and symptom entity and the corresponding context information in the target medical record text.

[0089] In one example, if Figure 5 As shown, the process is implemented through the following steps S1031-S1034:

[0090] S1031, using the entity recognition model to predict the target medical record text to obtain disease entities and symptom entities.

[0091] S1032, using a time period segmentation model to segment the target medical record text, and taking the beginning part and the middle part of the predicted continuous time period as the text content in one time period; the target medical record text is composed of texts of different time periods.

[0092] S1033, searching the target medical record text for the disease entity and the symptom entity; wherein the disease entity and the symptom entity exist in the corresponding text respectively.

[0093] S1034, taking the disease entity and its corresponding text and the symptom entity and its corresponding text as text pairs, as the disease entity and its corresponding context information and the symptom entity and its corresponding context information respectively.

[0094] S20, based on the disease entity and symptom entity information and corresponding context information, determine the masculine and feminine nature of the disease entity and symptom entity in the target medical record text.

[0095] In one example, if Figure 6 As shown, the process is implemented through the following steps S201-S203:

[0096] S201, annotating the disease entity and its corresponding context information and the symptom entity and its corresponding context information, obtaining the positive and negative information of the disease entity and the symptom entity, and constructing training data for judging the positive and negative of the disease entity and the symptom entity; the annotated labels include positive, negative, and cannot be judged.

[0097] S202, using the medical record text to obtain a pre-trained language model.

[0098] S203, using the pre-trained language model, fine-tuning the constructed training data to obtain the positive and negative results of the disease entities and symptom entities in the target medical record text.

[0099] In one example, the process is implemented by the following steps:

[0100] The labeled training data is concatenated from w and p to form the input U = [u1,…,u n ] and add the label [CLS] before the sequence input from the source end; where w and p represent entity and context information respectively;

[0101] Define Trm() as BERT bio The encoding calculation unit in the code, each word passes through the encoding unit to obtain the encoding representation, which is calculated by the following formula:

[0102]

[0103] Among them, u t is each item of the input vector U, and trm is "BERT bio The encoding calculation unit in is the result of coding calculation.

[0104] BERT bio Take the hidden layer vector representation of the first [CLS] of the model After a layer of fully connected linear transformation:

[0105]

[0106] The probability distribution of the target medical text is given by softmax:

[0107] Prob = softmax(C)

[0108] The maximum probability is taken as the positive or negative identification result of the entity.

[0109] The following is a detailed introduction to the above process through a specific implementation process, including steps 1 to 4:

[0110] Step 1: Use the medical record text with entity annotation information to train a sequence labeling model for entity recognition.

[0111] Step 1.1: Collect medical record text data, and have professional annotation personnel annotate the disease and symptom entities in the text.

[0112] Step 1.2: Convert the data obtained in step 1.1 into BIO annotation format.

[0113] Here x is the medical record text, x=[x1,…,x n ], x i represents each character; y is the labeling result, y=[y1,…,y n ],y i By B sym , I sym , B dis , I dis and O, which represent the beginning character of a symptom, the middle character of a symptom, the beginning character of a disease, the middle character of a disease, and the non-disease symptom character respectively.

[0114] Step 1.3: Train a sequence tagging model NER using the data in 1.2 bio , that is, the entity recognition model.

[0115] Step 1.3.1: Initialize the input vector using the word vectors obtained from large-scale medical text pre-training:

[0116] h i =f emb (x i )

[0117] Among them, f emb Indicates a method (f) of encoding (emb), h i Represents x i The result after f encoding is represented.

[0118] Step 1.3.2: Train a sequence tagging model NER using a bidirectional recurrent neural network based on LSTM units bio :

[0119] s i =BiLSTM bio (h i )

[0120] Step 2: Use large-scale medical record text to train a sequence labeling model for time segmentation, namely the time segmentation model:

[0121] Step 2.1: Annotators with professional knowledge divide the time periods in the medical record text.

[0122] Step 2.2: Convert the data obtained in step 2.1 into BIO annotation format. Here x is the medical record text, x = [x1,…,x n ], x i represents each character; y is the labeling result, y=[y1,…,y n ],y i By B part , I part and O, which represent the beginning character of a time period, the middle character of a time period, and the non-time period description character respectively.

[0123] Step 2.3: Train a sequence tagging model NER using the data in 2.2 split :

[0124] Step 2.3.1: Initialize the input vector using the word vectors pre-trained from medical text:

[0125] h i =f emb (x i )

[0126] Step 2.3.2: Train a sequence tagging model NER using a bidirectional recurrent neural network based on LSTM units split :

[0127] s i =BiLSTM split (hi )

[0128] Step 3: Use the model trained in steps 1 and 2 to construct training data for judging yin and yang.

[0129] Step 3.1: Use the model NER obtained in step 1 bio Predictions are made from the target medical record long text to obtain disease entities and symptom entities, using W dis and W sym express.

[0130] Step 3.2: Use step 2 to get the sequence tagging model NER split The target medical record text is segmented and the predicted continuous B part , I part As the text content in a time period. In this way, the target medical record text x is composed of texts in different time periods: x = [p1,…,p n ]; where p n This means that the target medical record text is divided into n parts (p), namely the first part (p1), the second part (p2), ..., the nth part (p n )

[0131] Step 3.3: Search the target medical record text for the disease entity W in step 3.1 dis and symptom entity W sym , disease entity W dis and symptom entity W sym Exist in the text p i and text p j .

[0132] Step 3.4: W dis and text p i , W sym and text p j As a text pair, it represents an entity and its corresponding contextual information, and is annotated by personnel with professional knowledge to obtain the positive and negative information of the symptom or disease entity. The annotated labels are PRESENT, ABSENT and UNKNOWN, which respectively represent positive, negative and cannot be judged.

[0133] Step 4: Use the language model pre-trained on large-scale medical text to fine-tune the training data constructed in step 3.

[0134] Step 4.1: The labeled training data is concatenated from w and p to form the input U = [u1,…,u n ], w and p represent entity and context information respectively, and the label [CLS] is added before the sequence input at the source end.

[0135] Step 4.2: Define Trm() for BERT bio The encoding calculation unit in the code, each word passes through the encoding unit to obtain the encoding representation, which can be calculated by the following formula:

[0136]

[0137] Among them, u t is each item of U in 4.1, and trm is "BERT bio The encoding calculation unit in is the result of coding calculation.

[0138] Step 4.3: For the task of distinguishing the masculine and feminine nature of entities, BERT bio Directly take the hidden layer vector representation of the first [CLS] of the model After a layer of fully connected linear transformation, it is as follows:

[0139]

[0140] Step 4.4: Give the probability distribution of the target medical text (correct or incorrect) through softmax:

[0141] Prob = softmax(C)

[0142] Step 4.5: Select the one with the highest probability as the discrimination result, and finally obtain the positive and negative identification result of the entity.

[0143] The method provided by the present invention has the following advantages:

[0144] It can avoid the problem of relying on templates and dictionaries, which are limited by the scale of templates and dictionaries and cannot handle undefined situations. At the same time, it can also continuously improve the accuracy of judgment by increasing the corpus and updating the models in the two stages of training.

[0145] Corresponding to the above-mentioned embodiment method, the present invention also provides a device for judging the positive and negative nature of disease entities and symptom entities in medical record texts. Figure 7 is a schematic diagram of the device structure, such as Figure 7 As shown, the device includes an acquisition module 1 and a judgment module 2, wherein:

[0146] The acquisition module 1 is used to obtain the disease entity and symptom entity and the corresponding context information in the target medical record text.

[0147] In one example, Figure 8 It is a schematic diagram of the structure of the acquisition module, such as Figure 8 As shown, the acquisition module 1 includes a first acquisition unit 11, a second acquisition unit 12 and a third acquisition unit 13:

[0148] The first acquisition unit 11 is used to obtain an entity recognition model.

[0149] The second acquisition unit 12 is used to obtain a time segmentation model.

[0150] The third acquisition unit 13 is used to obtain the disease entity and symptom entity and the corresponding context information in the target medical record text by using the entity recognition model and the time period segmentation model.

[0151] The judgment module 2 is used to judge the masculine and feminine nature of the disease entity and the symptom entity in the target medical record text based on the disease entity and symptom entity information and the corresponding context information.

[0152] The functions performed by the various components in the device provided in the embodiment of the present invention have been described in detail in the above method, so they will not be described in detail here.

[0153] Corresponding to the above embodiment, the embodiment of the present invention also provides a system for judging the positive and negative nature of disease entities and symptom entities in medical record texts, specifically as follows: Fig. 9 As shown, the system includes at least one processor 91 and a memory 92;

[0154] Memory 92, for storing one or more program instructions;

[0155] The processor 91 is used to run one or more program instructions to execute any method step in the method for determining the positive or negative nature of disease entities and symptom entities in the medical record text as described in the above embodiment.

[0156] Corresponding to the above-mentioned embodiment, an embodiment of the present invention further provides a chip, which is coupled to the memory in the above-mentioned system, so that the chip calls the program instructions stored in the memory during operation to implement the method for determining the positive and negative nature of disease entities and symptom entities in the medical record text as introduced in the above-mentioned embodiment.

[0157] Corresponding to the above-mentioned embodiments, an embodiment of the present invention further provides a computer storage medium, which includes one or more programs, wherein one or more program instructions are used by a system for judging the positive and negative nature of disease entities and symptom entities in medical record texts to execute the method for judging the positive and negative nature of disease entities and symptom entities in medical record texts as introduced above.

[0158] The solution provided in this application can more accurately identify the symptoms and disease information in the medical text through two stages, and make a positive and negative judgment that is consistent with the context.

[0159] The professionals should further realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to the function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0160] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be implemented using hardware, a software module executed by a processor, or a combination of the two. The software module may be placed in a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

[0161] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for determining the positive and negative nature of diseases and symptoms in medical records, characterized in that: The determination method comprises: Obtain an entity recognition model; Obtain the time segmentation model; Use the entity recognition model to predict the target medical record text to obtain disease entities and symptom entities; The target medical record text is segmented using a time segmentation model, and the beginning part and the middle part of the predicted continuous time segment are regarded as the text content in one time segment; the target medical record text is composed of texts of different time segments; Searching the target medical record text for the disease entity and the symptom entity; wherein the disease entity and the symptom entity exist in the corresponding text respectively; The disease entity and its corresponding text and the symptom entity and its corresponding text are taken as text pairs, namely, the disease entity and its corresponding context information and the symptom entity and its corresponding context information; The disease entity and its corresponding context information and the symptom entity and its corresponding context information are labeled to obtain the positive and negative information of the disease entity and the symptom entity, and to construct training data for judging the positive and negative of the disease entity and the symptom entity; the labeled labels include positive, negative, and cannot be judged; Use medical record text to obtain a pre-trained language model; The pre-trained language model is used to fine-tune the constructed training data to obtain the positive and negative results of disease entities and symptom entities in the target medical record text.

2. The determination method according to claim 1, characterized in that: The step of obtaining the entity recognition model comprises the following steps: Collect medical record texts, and annotate disease entities and symptom entities in the medical record texts; Convert the annotated medical record text data into BIO annotation format; Initialize the input vector with the word vectors obtained by pre-training the medical record text data in BIO annotation format; According to the initialized input vector, a bidirectional recurrent neural network training based on LSTM unit is used to obtain an entity recognition model.

3. The determination method according to claim 1, characterized in that: The step of obtaining the time segmentation model comprises the following steps: Divide the time periods in the medical record text; Convert the processed medical record text sample data into BIO annotation format; Initialize the input vector with the word vectors obtained by pre-training the medical record text data in BIO annotation format; According to the initialized input vector, a bidirectional recurrent neural network training based on LSTM unit is used to obtain a time segmentation model.

4. The determination method according to claim 1, characterized in that: The method of using the pre-trained language model to fine-tune the constructed training data to obtain the positive and negative results of the disease entity and the symptom entity in the target medical record text includes the following steps: The labeled training data is and Splicing input , and add the label [CLS] before the sequence input at the source end; and Represent entity and context information respectively; definition for The encoding calculation unit in the code, each word passes through the encoding unit to obtain the encoding representation, which is calculated by the following formula: , Where ut is each item of the input vector U; Take the hidden layer vector representation of the first [CLS] of the model , after a layer of fully connected linear transformation: , The probability distribution of the target medical text is given by softmax: , The maximum probability is taken as the positive or negative identification result of the entity.

5. A device for judging the positivity of disease entities and symptom entities in medical record texts, characterized in that: The judging device comprises: Get the module, Obtain an entity recognition model; Obtain the time segmentation model; Use the entity recognition model to predict the target medical record text to obtain disease entities and symptom entities; The target medical record text is segmented using a time segmentation model, and the beginning part and the middle part of the predicted continuous time segment are regarded as the text content in one time segment; the target medical record text is composed of texts of different time segments; Searching the target medical record text for the disease entity and the symptom entity; wherein the disease entity and the symptom entity exist in the corresponding text respectively; The disease entity and its corresponding text and the symptom entity and its corresponding text are taken as text pairs, namely, the disease entity and its corresponding context information and the symptom entity and its corresponding context information; Judgment module, The disease entity and its corresponding context information and the symptom entity and its corresponding context information are labeled to obtain the positive and negative information of the disease entity and the symptom entity, and to construct training data for judging the positive and negative of the disease entity and the symptom entity; the labeled labels include positive, negative, and cannot be judged; Use medical record text to obtain a pre-trained language model; The pre-trained language model is used to fine-tune the constructed training data to obtain the positive and negative results of disease entities and symptom entities in the target medical record text.

6. The judgment device according to claim 5, characterized in that: The acquisition module comprises: A first acquisition unit, used to obtain an entity recognition model; A second acquisition unit is used to obtain a time segmentation model; The third acquisition unit is used to obtain the disease entity and symptom entity and corresponding context information in the target medical record text by using the entity recognition model and the time period segmentation model.

7. A system for judging the positivity of disease entities and symptom entities in medical records, characterized in that: The system includes at least one processor and a memory; The memory is used to store one or more program instructions; The processor is configured to execute one or more program instructions to perform the method according to one or more of claims 1 to 4.

8. A chip, characterized in that: The chip is coupled to a memory in the system, so that the chip calls program instructions stored in the memory when running to implement one or more of the methods described in claims 1 to 4.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes one or more program instructions, which can be executed by the system according to claim 7 to implement the method according to one or more of claims 1 to 4.