A regional medical information data integration method, device and system

Abstract

The present invention relates to the field of text processing, and specifically provides a method, device and system for integrating regional medical information data. For a selected medical training text in a selected medical training text sub-library in a medical training text library, based on each text information recognition component of an initial text processing neural network, a first text information recognition error under a first pending text segment is determined, and a second text information recognition error under a second pending text segment of the selected medical training text is determined; the initial text processing neural network is adjusted according to the first text information recognition error and the second text information recognition error respectively corresponding to each medical training text in the medical training text library. The present invention can make the second pending text segment corresponding to the selected medical training text sub-library not involve the adjustment of text information recognition components other than the selected text information recognition component, thereby improving the accuracy of target medical information extraction from different data sources in the region during data integration.

Description

Technical Field

[0001] The present invention relates to the field of text processing, and in particular to a method, device and system for integrating regional medical information data. Background Art

[0002] With the continuous development of medical technology and the accumulation of medical data, information integration between different medical institutions in the region has become increasingly important. Integrating regional medical information helps to share medical resources, improve the accuracy of medical diagnosis, promote medical research, and improve the overall quality of medical services. However, there are many challenges in integrating regional medical information data. Medical information comes from a wide range of sources and is complex. Different medical institutions may use different information systems, and there is a lack of uniformity in data formats and standards.

[0003] In order to extract useful target medical information from a large amount of medical information text, text processing neural networks are widely used. Neural networks have strong learning capabilities and can automatically learn patterns and features from a large amount of text data to classify, identify and extract medical information. For example, neural networks can identify target medical information such as disease names, symptom descriptions, and treatment methods from medical record texts.

[0004] At present, there are some urgent problems to be solved when using text processing neural networks for medical information integration. In the process of neural network tuning, multiple training text libraries are usually used. However, there are often cases where the target medical information is inconsistent or the structure is different between these training text libraries. For example, negative samples in one training text library are identified as positive samples in other training text libraries. This inconsistency will interfere with the neural network learning process, resulting in the network being unable to fully utilize the information of all training samples. Summary of the invention

[0005] The purpose of the present invention is to provide a method, device and system for integrating regional medical information data.

[0006] The present invention is achieved in that:

[0007] In a first aspect, the present invention provides a method for integrating regional medical information data, the method comprising: obtaining target medical information text; calling a target text processing neural network, loading the target medical information text into the target text processing neural network, so as to identify the target medical information text based on the target text processing neural network and obtain target medical information; integrating the identified target medical information into a target space of a preset medical information library; wherein the target text processing neural network is obtained by the following steps: obtaining a medical training text library, the medical training text library comprising a plurality of medical training text sub-libraries, each of which corresponds to a target medical information with an error; for a selected medical training text sub-library among the plurality of medical training text sub-libraries, determining a selected text information recognition component set corresponding to the selected medical training text sub-library in each text information recognition component of the initial text processing neural network; wherein the selected text information recognition component in the selected text information recognition component set and the target medical information corresponding to the selected medical training text sub-library Correspondingly; for the selected medical training text in the selected medical training text sub-library, based on the various text information recognition components, according to the various first pending text segments corresponding to the selected medical training text, determine the first text information recognition error corresponding to the selected medical training text; wherein the first pending text segment is inferred to have target medical information corresponding to the medical training text; based only on the selected text information recognition components in the selected text information recognition component set, according to the various second pending text segments corresponding to the selected medical training text, determine the second text information recognition error corresponding to the selected medical training text; wherein the second pending text segment is inferred not to have target medical information corresponding to the medical training text; according to the first text information recognition error and the second text information recognition error respectively corresponding to the various medical training texts in the medical training text library, adjust the initial text processing neural network to obtain the target text processing neural network, and each text information recognition component in the target text processing neural network is configured to recognize the target medical information of the medical training text library.

[0008] In a second aspect, the present invention provides a regional medical information data integration device, the device comprising: a text acquisition module, used to acquire target medical information text; a network calling module, used to call a target text processing neural network, load the target medical information text into the target text processing neural network, and identify the target medical information text based on the target text processing neural network to obtain target medical information; an information integration module, used to integrate the identified target medical information into a target space of a preset medical information library; a network adjustment module, used to adjust the target text processing neural network using the following process:

[0009] Acquire a medical training text library, wherein the medical training text library includes a plurality of medical training text sub-libraries, and each of the medical training text sub-libraries respectively corresponds to target medical information with errors;

[0010] For a selected medical training text sub-library among the multiple medical training text sub-libraries, determining a selected text information recognition component set corresponding to the selected medical training text sub-library in each text information recognition component of the initial text processing neural network; wherein the selected text information recognition components in the selected text information recognition component set correspond to the target medical information corresponding to the selected medical training text sub-library;

[0011] For the selected medical training text in the selected medical training text sub-library, based on the respective text information recognition components, according to the respective first pending text segments corresponding to the selected medical training text, determine the first text information recognition error corresponding to the selected medical training text; wherein the first pending text segment is inferred to have target medical information corresponding to the medical training text;

[0012] Based only on the selected text information recognition components in the selected text information recognition component set, according to each second pending text segment corresponding to the selected medical training text, the second text information recognition error corresponding to the selected medical training text is determined; wherein the second pending text segment is inferred to have no target medical information corresponding to the medical training text;

[0013] The initial text processing neural network is calibrated based on the first text information recognition error and the second text information recognition error respectively corresponding to each medical training text in the medical training text library to obtain the target text processing neural network, and each text information recognition component in the target text processing neural network is configured to recognize the target medical information of the medical training text library.

[0014] In a third aspect, the present invention provides a medical information system comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, they implement the method described above.

[0015] The beneficial effects of the present invention are as follows: the method and system provided by the present invention adopts the method of determining the selected text information recognition component corresponding to the selected medical training text sub-library from among the various text information recognition components corresponding to the initial text processing neural network, and based on the selected text information recognition component, according to the second pending text segment corresponding to the medical training text in the selected medical training text sub-library, determining the text information recognition error of the second pending text segment under the selected text information recognition component, so that the second pending text segment corresponding to the selected medical training text sub-library does not involve the adjustment of text information recognition components other than the selected text information recognition component, and the second pending text segment corresponding to the selected medical training text sub-library only works in the selected medical training text sub-library, eliminating the network learning interference caused by the inconsistency of target medical information or structural differences between these medical training text libraries, and the inability to fully utilize the information of all training samples, resulting in the problem of limited adjustment performance, thereby improving the accuracy of target medical information extraction of different data sources in the region during data integration.

[0016] Furthermore, because the first pending text segment corresponding to the medical training text in the selected medical training text sub-library can be used for learning and adjustment of all text information recognition components corresponding to the initial text processing neural network, the negative features corresponding to the medical training text sub-library outside the selected medical training text sub-library are strengthened, making the information detection effect of the initial text processing neural network more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of a method for integrating regional medical information data provided by an embodiment of the present invention.

[0018] Figure 2 It is a flow chart of a calibration process of a target text processing neural network provided by an embodiment of the present invention.

[0019] Figure 3 It is a schematic diagram of the functional module architecture of the regional medical information data integration device provided in an embodiment of the present invention.

[0020] Figure 4 It is a schematic diagram of the composition of a medical information system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The execution subject of the regional medical information data integration method in the embodiment of the present invention is a medical information system, including but not limited to a server, a personal computer, a notebook computer, a tablet computer, a smart phone, etc. Figure 1 As shown, the method includes:

[0022] Step S100: Obtain target medical information text.

[0023] In a regional medical environment, medical information may be scattered in different medical institutions, such as hospitals, clinics, community health service centers, etc. Each medical institution has its own information recording system, which stores a variety of medical data. Among them, medical information in text form is the source of the target medical information text. For example, the hospital's electronic medical record system contains the patient's basic information (name, age, gender, etc.), symptom description (such as headache, fever, etc.), diagnosis results (such as cold, pneumonia, etc.), treatment process (drug use, surgical records, etc.), etc. These contents recorded in text form belong to the target medical information text. The clinic's diagnosis and treatment records, including brief text records such as the patient's chief complaint and preliminary diagnosis, are also part of the target medical information text. The text records of residents' health files in the community health service center, such as the condition monitoring data and health guidance suggestions in the chronic disease management records, are also within the scope of the target medical information text.

[0024] Specifically, the medical information system can obtain these target medical information texts in a variety of ways. A common way is to establish a data interface and connect with the information systems of various medical institutions. Taking the hospital's electronic medical record system as an example, if it uses a standard medical information exchange format (such as HL7-Clinical Document Architecture (CDA)), the medical information system can obtain the text information in it by parsing the interface in this format. Assume that a patient's medical record document in the hospital's electronic medical record system complies with the HL7-CDA standard, and its document structure contains multiple chapters, such as the patient information chapter, the medical history chapter, etc. The medical information system can extract the text content by parsing specific tags in this standard structure. For example, for the patient's symptom description part, it may exist in a file called <symptom-description>The medical information system can identify the tag and extract the text in it and incorporate it into the target medical information text.

[0025] In addition, for some medical information texts stored in non-standard formats, the medical information system can use data mining and text extraction technology. For example, some small clinics may use simple office software (such as Word documents) to record patients' diagnosis and treatment information. The medical information system can use text mining algorithms to first identify medical-related keywords (such as disease names, therapeutic drug names, etc.) in these documents, and then extract paragraphs or sentences containing these keywords as target medical information texts. The specific algorithm can be based on the term frequency-inverse document frequency (TF-IDF) algorithm to determine the importance of keywords. The formula of TF-IDF is: TF-IDF=TF*IDF, where TF (term frequency) represents the frequency of a word in a document, and the calculation formula is: TF=(the number of times the word appears in the document) / (the total number of words in the document); the calculation formula of IDF (inverse document frequency) is: IDF=log((total number of documents) / (the number of documents containing the word + 1)). By calculating the TF-IDF value of each word, the words with higher TF-IDF values ​​are selected as keywords, and then the text content containing these keywords is extracted.

[0026] It should be noted that in the process of obtaining the target medical information text, the security and legality of the data need to be considered. Medical information contains the patient's private information, such as the patient's identity information, medical condition privacy, etc. The medical information system must comply with relevant laws and regulations.

[0027] Step S200: Retrieve the target text processing neural network, load the target medical information text into the target text processing neural network, recognize the target medical information text based on the target text processing neural network, and obtain the target medical information.

[0028] The target text processing neural network is a pre-tuned neural network model specifically designed to process medical information text. This neural network model has a specific structure and parameters that can effectively analyze and process the input medical information text. For example, the neural network may contain multiple hidden layers, each of which is composed of multiple neurons, and the neurons are connected by specific weights. This structure enables the neural network to learn complex patterns in medical information text.

[0029] When the medical information system performs this step, it first calls the target text processing neural network. This calling process involves loading the neural network model from the location where the model is stored (such as a local disk, a specific model storage server, etc.) into the memory for subsequent computing operations. In the actual operating environment of the medical information system, there may be multiple neural network models for different tasks. The medical information system needs to accurately identify and call the target text processing neural network built specifically for medical information text processing.

[0030] After calling the target text processing neural network, the medical information system will load the previously acquired target medical information text into this neural network. This loading process is similar to feeding raw materials into a production and processing machine. For example, if the target medical information text is an electronic medical record about a patient, which contains information such as the patient's symptoms, medical history, and test results, then these text contents will be converted into a format that the neural network can process. A common way is to convert the text into a vector representation. A word vector model (such as Word2Vec) can be used to achieve this conversion. Assuming that there are words such as "headache" and "fever" in the target medical information text, the Word2Vec model will map these words into vectors of fixed dimensions (for example, 100 dimensions). These vectors constitute the vector representation of the target medical information text, which is then loaded into the target text processing neural network.

[0031] The process of identifying the target medical information text based on the target text processing neural network is to use the knowledge and patterns that have been learned in the neural network to parse the input medical information text. For example, the neural network may have learned the association pattern between certain symptom combinations and specific diseases in the previous training process. When the target medical information text containing these symptom information is input, the neural network will calculate the output of each neuron according to its internal weights and activation functions, thereby gradually identifying the target medical information in the text. Taking a simple neural network structure as an example, assume that the input layer has n neurons, corresponding to the n dimensions of the target medical information text vector, the hidden layer has m neurons, and the output layer has k neurons. For each neuron i (1≤i≤n) in the input layer, its output value is xi (that is, the i-th component of the target medical information text vector). The input of the neuron j (1≤j≤m) in the hidden layer is ∑(wi*xi) (where wi is the connection weight from the input layer neuron i to the hidden layer neuron j), and then the output yj of the hidden layer neuron j is obtained through the activation function (such as ReLU function: f(x)=max(0, x)). The input of the neuron l (1≤l≤k) in the output layer is ∑(wj*yj) (where wj is the connection weight from the hidden layer neuron j to the output layer neuron l). After the activation function (such as the Softmax function: Softmax(xi)=exp(xi) / ∑exp(xj)), the output of the output layer neuron l is obtained. This output can represent the probability distribution of different categories of target medical information (such as disease type, severity of the disease, etc.).

[0032] Through such a recognition process, the medical information system can obtain the target medical information from the target medical information text. For example, if the target medical information text is about the preliminary diagnosis of a patient, the neural network may identify key information, such as the name of the disease is "pneumonia", the severity of the disease is "mild", etc. These identified information are the target medical information.

[0033] Throughout the process, the medical information system needs to ensure the accuracy and reliability of the target text processing neural network. This may involve regular evaluation and updating of the neural network to adapt to the ever-changing medical information patterns and needs. At the same time, the medical information system also needs to deal with possible abnormal situations, such as the input target medical information text format does not meet the requirements, the neural network overfitting or underfitting and other problems. For situations where the format does not meet the requirements, the medical information system can set up a data preprocessing module to convert and clean the input text format; for the overfitting or underfitting problem of the neural network, cross-validation, adjustment of the neural network structure (such as increasing or decreasing the number of hidden layers, neurons, etc.), adjustment of regularization parameters and other technical means can be used to solve it.

[0034] Step S300: Integrate the identified target medical information into the target space of the preset medical information library.

[0035] The preset medical information database is a structure in the medical information system that is specifically used to store medical information. It has a certain organizational form and storage rules. This information database can be regarded as a large medical knowledge container, which contains medical information from different sources, which is classified and stored according to a specific logic. For example, it may be divided into different categories such as patient personal information, disease diagnosis information, and treatment process information, and each category may contain multiple subcategories. Taking the patient's disease diagnosis information as an example, the storage area may be further subdivided according to the type of disease (such as medical disease, surgical disease, etc.). The target space is a specific area in this medical information database that is pre-set and specifically used to store the current integrated target medical information. The setting of this target space helps to manage different types of medical information in an orderly manner and avoid confusion and incorrect storage of information.

[0036] When the medical information system performs this step, it first ensures that the target medical information obtained by identification is accurate and complete. In step S200, the target medical information is obtained by identifying the target medical information text through the target text processing neural network. For example, if the target medical information text is a medical record about a patient, the neural network identifies the patient's name as "Zhang San", age as "35 years old", diagnosis as "hypertension", treatment drug as "nifedipine" and other information. These identified information are the target medical information that needs to be integrated into the target space of the medical information library. When integrating the target medical information into the target space, the medical information system needs to adopt appropriate technical means to ensure the correct storage of the information. A common technical means is to use a database management system (DBMS), such as a relational database (such as MySQL, Oracle, etc.) or a non-relational database (such as MongoDB, etc.). If a relational database is used, the medical information system needs to design the corresponding table structure according to the content of the target medical information and the structure of the target space. For example, for the basic information of the patient, a table called "Patient Basic Information Table" can be created, which contains fields such as "Name", "Age", and "Gender". For the diagnosis result information, you can create a "diagnosis result table" containing fields such as "patient ID" (associated with the patient basic information table), "disease name", and "diagnosis time". The medical information system stores the identified target medical information according to the definition of these table structures. Assuming that the information of patient "Zhang San" is to be stored, in the "patient basic information table", an SQL statement similar to "INSERT INTO patient basic information table (name, age) VALUES ('Zhang San', 35)" will be executed (this is just a simple example, and the actual application may contain more fields and integrity constraints, etc.).

[0037] To ensure the integrity of data, the medical information system needs to handle possible missing or inconsistent information. For example, if the patient's gender information is not accurately identified during the process of identifying the target medical information, the medical information system cannot directly store the incomplete information in the target space. It can adopt some strategies, such as marking the information as missing, or making inferences based on other relevant information (if there is a reasonable basis for the inference). If there is inconsistent information, such as the same patient being identified as different ages in different medical record texts (possibly due to data entry errors, etc.), the medical information system needs to clean and correct the data. This inconsistency can be resolved by setting some rules, such as taking the latest and most reliable source of information as the final stored information.

[0038] In addition, when integrating the target medical information, data security and privacy protection also need to be considered. Medical information contains sensitive information of patients, such as medical condition privacy. The medical information system needs to comply with relevant laws, regulations and privacy policies to encrypt the information stored in the medical information database. For example, a symmetric encryption algorithm (such as the AES algorithm) can be used to encrypt and store sensitive information such as the patient's name and diagnosis results. The encryption process can be simply expressed as: let the plaintext be M (that is, the sensitive part of the target medical information), the key be K, and the encrypted ciphertext C=AES_Encrypt(M, K). When you need to query or use this information, the original information is restored through the decryption algorithm (C=AES_Decrypt(C, K)).

[0039] With the development of medical technology and the continuous increase of medical data, the medical information database also needs to be continuously optimized and expanded. The medical information system may need to regularly maintain the medical information database, including adjusting the table structure (such as adding new fields to adapt to new types of medical information), optimizing the index (improving query efficiency), and other operations. For example, with the increasing application of genetic testing information in medical diagnosis, the medical information system may need to add a structure specifically for storing information related to genetic testing results in the medical information database, and optimize the relevant query indexes so that the new medical information can be queried and used quickly and accurately.

[0040] Please refer to Figure 2 In the embodiment of the present invention, the target text processing neural network is obtained by the following steps:

[0041] Step S10: Acquire a medical training text library, where the medical training text library includes a plurality of medical training text sub-libraries, and errors exist in target medical information corresponding to each medical training text sub-library.

[0042] The medical training text library is composed of multiple medical training text sub-libraries. Each medical training text sub-library has its own unique significance in the entire neural network training system, which is due to the differences in the target medical information they correspond to. To illustrate with a specific example, suppose there is a medical training text library used to train a neural network for application in regional medical information integration. One of the training text sub-libraries focuses on disease diagnosis information. The text in this sub-library is mainly related text content such as the diagnosis basis and diagnosis results of various diseases. For example, there will be a large amount of text about disease diagnosis in this sub-library, such as text information that determines the disease through blood test indicators (such as the association between abnormal white blood cell count and infectious diseases), imaging examination results (such as the relationship between lung CT image features and pneumonia), etc. The target medical information here is the diagnosis-related information of the disease. When training this sub-library, the neural network mainly learns how to accurately identify these diagnosis-related medical information.

[0043] Another training text sub-library may focus on medical information related to treatment plans. The text in this sub-library contains treatment methods for different diseases, such as the type, dosage, and course of drug treatment, the operation process of surgical treatment, and precautions for postoperative recovery. For example, in the treatment plan sub-library for cancer, there will be text content about the selection of chemotherapy drugs (the effectiveness of different chemotherapy drugs for different types of cancer), the dosage and range setting of radiotherapy, and the applicable conditions of targeted therapy drugs. Its recognition target is various medical information related to the treatment plan.

[0044] When jointly training a neural network, due to the different target medical information in different sub-databases, a false value training text in one sub-database may be a true value training text in another sub-database. For example, in the disease diagnosis sub-database, if a specific symptom description is irrelevant to the target disease, it is a false value training text. However, in the symptom description sub-database (another sub-database), this symptom description itself is the target information to be identified (true value training text). Suppose that in the disease diagnosis sub-database, "mild cough" for pneumonia diagnosis is marked as a false value training text (because a single mild cough is not enough to diagnose pneumonia), but in the symptom description sub-database, "mild cough" is the symptom information to be accurately identified (true value training text). If this situation is not handled properly during the training process, it will cause confusion to the neural network.

[0045] The second is the problem of interference from text information recognition components. Neural networks contain multiple text information recognition components, similar to different parts of a classifier. If all text information recognition components participate in the training of all training text sub-libraries, problems will arise. For example, for the training of the disease diagnosis sub-library, some text information recognition components are specifically optimized for diagnosis-related features. But when these components participate in the training of the treatment plan sub-library, these components may be updated incorrectly because the text features in the treatment plan sub-library are very different from those in the diagnosis sub-library. It is like a component that is good at recognizing the structure and semantics of disease diagnosis-related text. When faced with completely different semantics and structures in the treatment plan text (such as drug dosage, surgical steps, etc.), if it participates in training, it may adjust its own parameters due to irrelevant information, resulting in a decrease in the original recognition ability of the diagnosis text, which in turn affects the performance of the entire neural network.

[0046] When obtaining such a medical training text library, the medical information system needs to carefully screen and organize it from a large number of medical data sources. For the disease diagnosis information sub-library, the medical information system may filter out text content containing relevant fields such as diagnosis results and diagnosis basis from the hospital's electronic medical record system. At the same time, in order to ensure the accuracy and completeness of the data, these texts may need to be pre-processed, such as standardizing some vague diagnostic terms. For the treatment plan sub-library, in addition to obtaining treatment-related information from the electronic medical record, it is also possible to obtain more detailed and standardized treatment plan text content from the hospital's treatment guide database.

[0047] When building these sub-databases, we also need to consider how to label the data so that the neural network can correctly distinguish different recognition targets. For the true value training text in the disease diagnosis sub-database, the key features related to the diagnosis of a specific disease should be clearly marked. For example, for the true value training text of pneumonia diagnosis, "lung shadows" (imaging features), "increased white blood cell count" (blood test features), etc. are marked as key true value training text features. For the true value training text in the treatment plan sub-database, such as a specific drug treatment plan for a certain disease, key information such as the drug name, dosage range, and applicable population should be marked.

[0048] Step S20: For a selected medical training text sub-library among multiple medical training text sub-libraries, determine a selected text information recognition component set corresponding to the selected medical training text sub-library in each text information recognition component of the initial text processing neural network; wherein the selected text information recognition component in the selected text information recognition component set corresponds to the target medical information corresponding to the selected medical training text sub-library.

[0049] The initial text processing neural network is a basic neural network structure that has multiple text information recognition components. These text information recognition components are similar to different processing units in a neural network, and each component plays a role in recognizing different types of text information. For example, these components may be divided into components for recognizing symptom information, components for recognizing disease names, components for recognizing treatment methods, etc. according to different functions.

[0050] When a medical information system processes multiple medical training text sub-databases, it is necessary to determine the corresponding selected text information recognition component set for the selected medical training text sub-database. Assume that there is a medical training text sub-database specifically for cardiovascular disease-related information, which contains text content of cardiovascular disease patients' symptoms (such as chest pain, palpitations, etc.), diagnosis results (such as coronary heart disease, arrhythmia, etc.) and treatment methods (such as drug treatment, interventional treatment, etc.).

[0051] For this selected cardiovascular disease medical training text sub-database, the medical information system needs to determine the appropriate component set among the various text information recognition components of the initial text processing neural network. If there is a component in the initial text processing neural network that is specifically used to recognize cardiovascular disease symptoms, then this component is likely to be selected into the selected text information recognition component set because it corresponds to the symptom-related text information in the cardiovascular disease medical training text sub-database. Similarly, if there is a component that is good at recognizing disease names and can effectively recognize cardiovascular disease names, it will also be selected into the component set.

[0052] Specifically, the medical information system can determine the selected text information recognition component set through a predefined mapping relationship. This mapping relationship can be based on the pre-analysis and marking of the function of each text information recognition component. For example, the medical information system can establish a table in which the columns represent the various text information recognition components and the rows represent different types of medical information (such as symptoms, disease names, treatment methods, etc.). In this table, if a text information recognition component is marked as being able to effectively process cardiovascular disease-related symptoms, then this component will be selected when processing the cardiovascular disease medical training text sub-library.

[0053] Assume that another selected medical training text sub-database is about endocrine diseases, which contains relevant text information about endocrine diseases such as thyroid disease and diabetes, including changes in hormone levels, disease symptoms, therapeutic drugs, etc. For this sub-database, the medical information system will again look for the corresponding text information recognition component in the initial text processing neural network. For example, there may be a component that has a good recognition ability for hormone-related terms and values. This component will be selected into the selected text information recognition component set for the endocrine disease medical training text sub-database because it corresponds to the target medical information such as changes in hormone levels in the sub-database.

[0054] This process of determining the set of selected text information recognition components is crucial for neural network training. If such a selection is not made, problems may occur during the training process. For example, if all text information recognition components are involved in the training of the cardiovascular disease medical training text sub-library, components that are not related to cardiovascular disease (such as components that specifically recognize orthopedic disease-related terms) may interfere with the training process. These irrelevant components may make incorrect adjustments due to receiving irrelevant text information, thereby affecting the training effect of the entire neural network.

[0055] At the same time, by determining the selected text information recognition component set, the neural network can be made more focused on the recognition of specific types of medical information. For example, in the training of the cardiovascular disease medical training text sub-library, the selected text information recognition component set can more intensively learn the text features related to cardiovascular disease, such as the combination relationship between symptoms, the association between diagnostic results and symptoms, etc. This helps to improve the accuracy and efficiency of the neural network in processing medical information related to cardiovascular disease.

[0056] In actual operation, the medical information system may need to conduct some evaluation and analysis of the initial text processing neural network to accurately determine the function and scope of application of each text information recognition component. This can be achieved by testing some known medical text samples. For example, a set of standard medical text samples containing information about various diseases are input into the initial text processing neural network respectively, and the output results of each text information recognition component are observed to determine which components have better recognition capabilities for which types of medical information.

[0057] In addition, as medical knowledge is constantly updated and medical data is continuously enriched, the medical information system may need to regularly adjust this mapping relationship and the selected text information recognition component set. For example, as new research results on cardiovascular diseases emerge, some new symptoms or diagnostic indicators may be discovered. At this time, it is necessary to check whether there are components in the initial text processing neural network that can process this new information. If not, it may be necessary to improve the components or re-determine the selected text information recognition component set.

[0058] Step S30: For the selected medical training text in the selected medical training text sub-library, based on each text information recognition component, according to each first pending text segment corresponding to the selected medical training text, determine the first text information recognition error corresponding to the selected medical training text; wherein, the first pending text segment is inferred to have target medical information corresponding to the medical training text.

[0059] In the neural network training process of regional medical information data integration, this step is an important link in evaluating the accuracy of the neural network in recognizing text information in the selected medical training text database, and the executor is the medical information system.

[0060] The selected medical training text sub-library is a part of the entire medical training text library, which focuses on specific target medical information. For example, this sub-library may be a medical information library specifically for a specific disease (such as diabetes). The selected medical training texts are the specific text content selected from this sub-library for training the neural network. These texts contain various information related to the target medical information, such as the patient's symptom description, diagnosis results, treatment process, etc.

[0061] The first pending text fragment is a part of the text content determined from the selected medical training text, which is closely related to the target medical information. Taking the diabetes medical training text sub-database as an example, if a selected medical training text describes the condition of a diabetic patient, the symptom description of "polydipsia, polyphagia, polyuria and weight loss" may be the first pending text fragment, because these symptoms are closely related to diabetes (target medical information).

[0062] Each text information recognition component is a component of the initial text processing neural network, and each component has its own specific recognition function, such as some components are good at recognizing symptom information, some are good at recognizing disease names, etc. In this step, the medical information system determines the first text information recognition error based on these text information recognition components. This error reflects the accuracy of the neural network in recognizing the target medical information contained in the first undetermined text segment in the selected medical training text.

[0063] Specifically, the medical information system can determine the error by comparing the recognition result of the text information recognition component on the first undetermined text segment with the actual target medical information. For example, assuming that the task of a certain text information recognition component is to recognize disease symptoms, for the first undetermined text segment of the above-mentioned diabetic patient "polydipsia, polyphagia, polyuria and weight loss", it may output a recognition result vector. If this vector differs from the exact representation of these symptoms (which can be based on a pre-defined standard encoding or feature vector), this difference will be quantified as part of the recognition error.

[0064] In order to calculate this error more accurately, some distance measurement methods can be used. For example, in the vector space, if the recognition result vector is , the standard target medical information representation vector is , we can use the Euclidean distance formula To measure the distance between the two, this distance can be used as a measure of recognition error. Here n is the dimension of the vector, and are the values ​​of the i-th dimension of the recognition result vector and the target medical information representation vector respectively.

[0065] If the recognition error is large, it means that the neural network has a large deviation in identifying text segments that are closely related to the target medical information, and the parameters of the neural network need to be adjusted to improve accuracy. For example, in a large-scale medical training process, if it is found that the recognition errors of various text information recognition components are generally large for the first pending text segments of many selected medical training texts, it may mean that there are problems with the structure or weight initialization of the neural network, and it is necessary to re-examine the construction of the neural network or adjust the training strategy.

[0066] In addition, different text information recognition components may have different contributions in determining the first text information recognition error. For example, for a selected medical training text containing multiple target medical information (such as symptoms, diagnosis results, etc.), the text information recognition component that recognizes symptoms and the component that recognizes diagnosis results may have different performances on different first pending text segments. The medical information system needs to comprehensively consider the error contribution of each component in order to accurately evaluate the recognition effect of the entire neural network on the selected medical training text, provide an accurate basis for subsequent neural network calibration, and thus improve the accuracy and reliability of target medical information recognition in the process of regional medical information data integration.

[0067] Step S40: Based only on the selected text information recognition component in the selected text information recognition component set, determine the second text information recognition error corresponding to the selected medical training text according to each second pending text segment corresponding to the selected medical training text; wherein, the second pending text segment is inferred to have no target medical information corresponding to the medical training text.

[0068] The selected text information recognition component set is a specific set of components that are previously determined from the various text information recognition components of the initial text processing neural network and correspond to the selected medical training text sub-database. These components are optimized for the target medical information in the selected medical training text sub-database. For example, if the selected medical training text sub-database is about cardiovascular diseases, then the selected text information recognition component set may include text information recognition components that are good at identifying cardiovascular disease symptoms (such as chest pain, palpitations, etc.) and cardiovascular disease diagnoses (such as coronary heart disease, myocardial infarction, etc.).

[0069] The selected medical training text is the text content selected from the selected medical training text sub-library for training the neural network. It contains a variety of information, some of which are related to the target medical information (such as the information related to the first pending text segment mentioned above), and the other part is not related to the target medical information, which involves the second pending text segment.

[0070] The second pending text segment is the part of the selected medical training text that does not have the target medical information corresponding to the medical training text. Take a selected medical training text about cardiovascular disease as an example. It may describe some other conditions of the patient, such as the patient's dietary preferences during hospitalization (prefer light food). This description of dietary preferences is the second pending text segment. Because it has no direct connection with cardiovascular disease (target medical information), it is impossible to directly infer information related to cardiovascular disease from this segment.

[0071] When the medical information system executes step S40, it only uses the components in the selected text information recognition component set to determine the second text information recognition error. The purpose of this operation is to avoid unnecessary interference of irrelevant text information recognition components on text segments unrelated to the target medical information, and ensure the accuracy and pertinence of the error calculation.

[0072] Specifically, the medical information system can determine this error in a variety of ways. A common way is to measure the recognition error by probability estimation. Assume that each selected text information recognition component is able to output a probability value for the input text segment as to whether it belongs to the target medical information. For example, for the second pending text segment of the above-mentioned dietary preferences, the selected text information recognition component for identifying cardiovascular disease symptoms may output a very low probability value, indicating that this segment is unlikely to be symptom information related to cardiovascular disease.

[0073] Suppose the probability value output by the selected text information recognition component i for the second pending text segment j is P ij , and we know in advance that this segment does not belong to the target medical information (based on prior knowledge or annotation), then we can define an error function. For example, we can use the logarithmic loss function To calculate the second text information recognition error. Here, all the second undetermined text segments j are summed. The principle of this function is that when the probability value P ij The closer it is to 0 (indicating that it is less likely to be the target medical information), The closer it is to 0, the smaller the error; conversely, if the probability value is high (incorrectly believed to be the target medical information), the error will increase.

[0074] This method of calculating the second text information recognition error based only on a selected set of text information recognition components helps improve the efficiency and accuracy of neural network training. Without such restrictions, all text information recognition components are involved in the calculation, which may lead to confusion in error calculation. For example, text information recognition components that are not related to cardiovascular disease (such as components that recognize information related to orthopedic diseases) may produce unreasonable outputs for the second pending text segment in the cardiovascular disease training text, thereby interfering with the error calculation and making it impossible for the neural network to accurately adjust according to the error.

[0075] In actual medical information data integration scenarios, accurately determining the second text information recognition error is crucial for the calibration of the neural network. For example, in a large training system containing a medical training text library of multiple disease types, if the second text information recognition error cannot be correctly calculated, it may cause the neural network to be confused when distinguishing text fragments related to and unrelated to the target medical information. This may further affect the neural network's ability to accurately recognize the target medical information, and ultimately affect the effective integration and utilization of medical information in the entire regional medical information data integration process.

[0076] In addition, different selected text information recognition components may have different contributions in determining the second text information recognition error. For example, in the set of selected text information recognition components for cardiovascular diseases, the component that recognizes cardiovascular disease diagnosis may play a more important role in determining whether certain second pending text fragments (such as descriptions related to other disease diagnoses) belong to the target medical information of cardiovascular disease, and its output probability value may have a relatively high weight when calculating the error. The medical information system needs to comprehensively consider these factors and reasonably determine the second text information recognition error so that in the subsequent neural network calibration process, the parameters of the selected text information recognition component can be accurately adjusted to improve the neural network's ability to recognize the target medical information, thereby better serving the regional medical information data integration task.

[0077] Step S50: According to the first text information recognition error and the second text information recognition error respectively corresponding to each medical training text in the medical training text library, the initial text processing neural network is adjusted to obtain the target text processing neural network, and each text information recognition component in the target text processing neural network is configured to recognize the target medical information of the medical training text library.

[0078] In the process of neural network construction for regional medical information data integration, this step is a key operation performed by the medical information system, which aims to optimize the initial text processing neural network based on the previously calculated error, so as to obtain a target text processing neural network that can accurately identify the target medical information of the medical training text library.

[0079] The medical training text library contains multiple medical training texts. For each medical training text, the corresponding first text information recognition error and second text information recognition error have been determined in the previous step. The first text information recognition error reflects the degree of deviation of the neural network when recognizing the text segment related to the target medical information (the first undetermined text segment), and the second text information recognition error reflects the error when processing the text segment unrelated to the target medical information (the second undetermined text segment).

[0080] For example, suppose there is a medical training text sub-library about respiratory diseases in the medical training text library, and one of the medical training texts describes the condition of a pneumonia patient in detail. In this text, the symptom description (such as cough, fever, shortness of breath, etc.) corresponds to the first pending text segment, while the description of the ward environment during the patient's hospitalization (such as ward orientation, bed number, etc.) belongs to the second pending text segment. For the text information recognition component that recognizes information related to respiratory diseases, if the recognition result of the symptom description part deviates from the actual situation, a first text information recognition error will occur; if the ward environment part is mistakenly judged to be related to respiratory diseases, a second text information recognition error will occur.

[0081] The medical information system uses these errors to adjust the initial text processing neural network. Specifically, this adjustment process involves a variety of methods. A common method is based on the gradient descent algorithm. Assuming that the parameter of the neural network is θ, the loss function J(θ) comprehensively considers the first text information recognition error and the second text information recognition error. The error for each medical training text can be regarded as a contribution to the loss function.

[0082] Taking the first text information recognition error as an example, if the mean square error (MSE) is used to measure this error, let the predicted value be , the actual value is , then the mean square error calculation formula of the first text information recognition error is , where n is the number of samples (here it can be understood as the number of the first undetermined text segments). For the second text information recognition error, a suitable error metric function can be defined in a similar way, and then the two are combined into the loss function L(θ).

[0083] In the gradient descent algorithm, the update formula for the parameter θ is , where α is the learning rate, is the gradient of the loss function L(θ) with respect to the parameter θ. By continuously updating the parameters according to this formula, the loss function L(θ) is gradually reduced, that is, the first text information recognition error and the second text information recognition error are gradually reduced. In this process, the parameters of different text information recognition components (such as components for recognizing symptoms, components for recognizing disease names, etc.) in the neural network will be adjusted according to their respective contributions to the error. For example, if the component for recognizing symptoms produces a large first text information recognition error in multiple medical training texts, then during the calibration process, the parameters corresponding to this component will be adjusted to a greater extent to improve its recognition accuracy of symptom information.

[0084] The importance of this calibration is that it enables the target text processing neural network to better adapt to the various text situations in the medical training text library. For medical information systems, accurately calibrating the neural network is the key to improving the efficiency and accuracy of regional medical information data integration. For example, in a large regional medical information integration project, the medical training text library may contain text information from different medical institutions covering various types of diseases. If the neural network is not accurately calibrated, a large number of errors may occur when identifying the target medical information, such as misjudging irrelevant information as important medical information, or failing to accurately identify key medical information.

[0085] By adjusting the first text information recognition error and the second text information recognition error corresponding to each medical training text, each text information recognition component in the target text processing neural network is reconfigured. After adjustment, these components can more accurately identify the target medical information of the medical training text library. For example, for a complex medical training text containing multiple disease diagnoses, symptoms and treatment information, the adjusted neural network can accurately extract various target medical information from the text, such as correctly identifying the name of the disease, accurately judging the relationship between symptoms and diseases, and reasonably analyzing treatment plans.

[0086] In addition, the calibration process also considers how to balance the impact of the first text information recognition error and the second text information recognition error. If too much attention is paid to the first text information recognition error, it may cause the neural network to overfit when processing text fragments that are not related to the target medical information, that is, it will also produce unnecessary responses to irrelevant information; conversely, if too much emphasis is placed on the second text information recognition error, it may affect the accurate recognition of the target medical information. The medical information system needs to reasonably adjust the weights of the two errors in the loss function according to the actual situation to achieve the best calibration effect.

[0087] As an implementation manner, step S40, determining the second text information recognition error corresponding to the selected medical training text based only on the selected text information recognition component in the selected text information recognition component set and according to each second pending text segment corresponding to the selected medical training text, may include:

[0088] Step S41: respectively determining the priori marks corresponding to the second pending text segments as first priori marks, wherein the first priori marks are used to indicate that the inferred medical information corresponding to the second pending text segment is not the target medical information corresponding to the selected medical training text sub-library;

[0089] Step S42: for the first selected text information recognition component in the selected text information recognition component set, determine the inferred medical information of each second to-be-determined text segment under the first selected text information recognition component according to the to-be-determined text semantic vectors corresponding to each second to-be-determined text segment; wherein the to-be-determined text semantic vectors are used to indicate the inferred medical information and the inferred medical information position corresponding to the to-be-determined text segment;

[0090] Step S43: determining the text information recognition error of each second undetermined text segment under the first selected text information recognition component according to the inferred medical information and the first priori mark of each second undetermined text segment under the first selected text information recognition component;

[0091] Step S44: taking the sum of the text information recognition errors of each second undetermined text segment under the first selected text information recognition component as the text information recognition error corresponding to the first selected text information recognition component;

[0092] Step S45: Determine a second text information recognition error corresponding to the selected medical training text according to the text information recognition errors respectively corresponding to the selected text information recognition components in the selected text information recognition component set.

[0093] A priori tag is a pre-set identifier used to indicate the relationship between a text segment and target medical information. For the second pending text segment, the medical information system will determine the first priori tag for it. The significance of this tag is to make it clear that the information contained in these segments is not the target medical information that the selected medical training text sub-library is concerned with. For example, assuming that the selected medical training text sub-library is about cardiovascular disease, its target medical information is information directly related to the diagnosis, symptoms, and treatment of cardiovascular disease. In a selected medical training text, there is a second pending text segment that describes the patient's dietary preferences during hospitalization (such as a preference for light food). The medical information system will determine the priori tag corresponding to this segment as the first priori tag, indicating that the target medical information related to cardiovascular disease cannot be inferred from this segment. The determination of this priori tag helps the subsequent recognition error calculation and provides a judgment benchmark for the medical information system.

[0094] Specifically, the medical information system can determine the first priori marker through predefined rules or a labeling system. When constructing a medical training text library, a labeling database can be established at the same time, which stores the relationship labels between each text segment and the target medical information. When processing the second pending text segment, the medical information system quickly determines its corresponding first priori marker based on the rules in this labeling database.

[0095] In step S42, the undetermined text semantic vector is a mathematical representation of the semantic information of the text segment, which includes not only the inferred medical information that the text segment may contain, but also the positional relationship of this information at the semantic level. In this step, the medical information system uses the undetermined text semantic vectors corresponding to each second undetermined text segment to determine the inferred medical information of each segment under this component for a specific first selected text information recognition component in the selected text information recognition component set.

[0096] Take a medical training text sub-database about diabetes as an example. The selected text information recognition component set has a first selected text information recognition component that specifically recognizes information related to diabetes complications. Assume that a selected medical training text has multiple second pending text segments, such as a segment describing the patient's daily exercise habits (walking for half an hour every day) and a segment describing other diseases in the patient's family (such as a family member with hypertension). These segments have their own corresponding pending text semantic vectors.

[0097] For a segment describing daily exercise habits, its pending text semantic vector may indicate that this segment has no direct association with diabetic complications, but has a certain semantic connection with health management. For a segment describing family disease conditions, its pending text semantic vector may indicate that this segment has little to do with diabetic complications, but is related to family medical history. By analyzing these pending text semantic vectors, under the first selected text information recognition component (recognizing information related to diabetic complications), the medical information system can determine the inferred medical information of each of the two second pending text segments, that is, neither of the two segments contains information related to diabetic complications.

[0098] Specifically, the semantic vector of the pending text can be constructed by combining word vector models with semantic role labeling and other technical means. For example, the words in the text fragment are converted into vectors using word vector models such as Word2Vec, and then the semantic relationship between the words is determined through semantic role labeling, thereby constructing the semantic vector of the pending text that can represent the inferred medical information and its positional relationship.

[0099] In step S43, at this step, the medical information system calculates the text information recognition error of each segment under this component based on the inferred medical information of each second pending text segment previously determined under the first selected text information recognition component and the pre-set first priori label.

[0100] Continuing with the above diabetes medical training text sub-library as an example, the first selected text information recognition component is to recognize information related to diabetes complications. For the second undetermined text segment describing daily exercise habits, its inferred medical information is not related to diabetes complications (according to step S42), and the first priori mark also indicates that this segment should not contain information related to diabetes complications (according to step S41). If the first selected text information recognition component mistakenly determines that this segment is related to diabetes complications, a recognition error will occur.

[0101] Assume that a simple binary judgment is used to calculate the error. If the judgment is correct (the inferred medical information is consistent with the first prior label), the error is 0; if the judgment is wrong, the error is 1. This simple calculation method may use more complex measurement methods in practice according to specific circumstances, such as probability-based error measurement. For example, if the first selected text information recognition component outputs that the probability of this segment being related to diabetic complications is p, and it should actually be 0 (according to the first prior label), then the logarithmic loss function L = -log(1-p) can be used to calculate the text information recognition error of this segment under this component.

[0102] In step S44, after calculating the text information recognition error of each second pending text segment under the first selected text information recognition component, the medical information system sums up these errors to obtain the text information recognition error corresponding to the selected component.

[0103] For example, in the diabetes medical training text sub-database, there are three second pending text segments T1, T2, and T3 whose text information recognition errors under the first selected text information recognition component (recognizing information related to diabetes complications) are e1=0.1, e2=0, and e3=0.2 respectively (here it is assumed that a certain probability-based error measurement method is used). Then the text information recognition error corresponding to the first selected text information recognition component is E=e1+e2+e3=0.1+0+0.2=0.3.

[0104] This summation result reflects the overall recognition error of the first selected text information recognition component when processing all the second pending text segments. It combines the error contributions of each segment and provides a quantitative indicator for evaluating the performance of this component in recognizing text segments that are not related to the target medical information.

[0105] In step S45, the medical information system determines the second text information recognition error corresponding to the entire selected medical training text according to the text information recognition error corresponding to each selected text information recognition component in the selected text information recognition component set. Assume that there are three components C1, C2, and C3 in the selected text information recognition component set of the diabetes medical training text sub-library, and their corresponding text information recognition errors are E1=0.3, E2=0.1, and E3=0.2, respectively. The medical information system can determine the second text information recognition error corresponding to the selected medical training text by weighted summation or other comprehensive methods.

[0106] For example, if a simple average weighted method is used, the second text information recognition error corresponding to the medical training text is selected . This second text information recognition error integrates the error conditions of each component in the selected text information recognition component set when processing the second pending text segment, and reflects the overall error level of the entire selected medical training text when the part unrelated to the target medical information is recognized. This error indicator is of great significance for the subsequent adjustment of the initial text processing neural network, and can help the medical information system adjust the parameters of the neural network to make it more accurate when processing text segments unrelated to the target medical information, avoid unnecessary interference, and thus improve the performance of the entire neural network in the embodiment of the present invention.

[0107] As an implementation mode, step S20, determining the selected text information recognition component set corresponding to the selected medical training text sub-library among the text information recognition components of the initial text processing neural network, may include:

[0108] Step S21: obtaining the text information recognition component activation state indicator corresponding to the selected medical training text sub-library, the text information recognition component activation state indicator being used to determine the activation state of each text information recognition component of the initial text processing neural network during text information recognition error calculation;

[0109] Step S22: determining a selected text information recognition component set from among the text information recognition components of the initial text processing neural network based on the text information recognition component activation state indicator;

[0110] Among them, the text information recognition component activation status indicator is determined based on the target medical information corresponding to the medical training text library and the distribution of text information recognition components corresponding to the initial text processing neural network.

[0111] In step S21, in the complex system of regional medical information data integration and neural network training, the text information recognition component activation status indicator is a key element. This indicator is obtained by the medical information system, and its purpose is to provide a basis for determining the activation status of each text information recognition component in a specific task (i.e., calculating the text information recognition error for the selected medical training text sub-library).

[0112] For example, suppose there is a medical training text sub-database that focuses on medical information related to cardiovascular disease, such as text content on cardiovascular disease symptoms (chest pain, palpitations, etc.), diagnosis (such as the basis for the diagnosis of coronary heart disease and arrhythmia), and treatment (drug therapy, interventional treatment plans, etc.). The initial text processing neural network contains multiple text information recognition components, which have different functions. Some may be good at recognizing symptom-related information, while others focus on diagnosis or treatment-related text recognition.

[0113] For this cardiovascular disease medical training text sub-library, there will be a corresponding text information recognition component activation status indicator. This indicator may be pre-set based on the previous analysis of the medical training text sub-library and the structure of the neural network. For example, if a text information recognition component is marked as mainly for bone disease related text recognition, then in the context of the cardiovascular disease medical training text sub-library, its corresponding activation status indicator may indicate that the component is in an inactive state when performing error calculations for this sub-library.

[0114] Specifically, the medical information system can obtain the activation status indicator of the text information recognition component through a configuration file or database. This configuration file or database can be created when building the association relationship between the medical training text library and the neural network. For example, a database in the form of a table can be established, with rows representing different medical training text sub-libraries, columns representing various text information recognition components in the initial text processing neural network, and each cell storing a corresponding activation status indicator (such as 0 for inactive and 1 for active). In this way, when processing the cardiovascular disease medical training text sub-library, the medical information system can directly query this database to obtain the activation status indicator corresponding to each text information recognition component.

[0115] In step S22, the medical information system determines the selected text information recognition component set from the numerous text information recognition components of the initial text processing neural network based on the text information recognition component activation status indicator obtained in step S21. This process is very critical because it is directly related to the ability of the neural network to effectively train the selected medical training text library.

[0116] Continuing with the example of the cardiovascular disease medical training text sub-database, the distribution of text information recognition components in the initial text processing neural network is diverse, covering a variety of possible medical information recognition functions. Based on the activation status indicator obtained previously, if the activation status indicator of a text information recognition component indicates that it should be activated when targeting the cardiovascular disease medical training text sub-database, then this component will be selected into the selected text information recognition component set.

[0117] Assume that there is a component in the initial text processing neural network that is specifically used to identify typical symptoms of cardiovascular disease, and its activation status indicator is 1 (indicating activation), because the function of this component matches the target medical information (including cardiovascular disease symptoms, etc.) of the cardiovascular disease medical training text sub-library. Therefore, this component will be selected into the selected text information recognition component set. If there is a component that is mainly used to recognize text related to ophthalmic diseases, its activation status indicator may be 0 (indicating inactivation), then this component will not be selected into the selected text information recognition component set for the cardiovascular disease medical training text sub-library.

[0118] From the perspective of technical implementation, this determination process can be achieved through simple conditional judgment. The medical information system traverses each text information recognition component of the initial text processing neural network and checks its corresponding activation status indicator. If the indicator meets the activation condition (for example, in the above example, the indicator is 1), the component is added to the selected text information recognition component set. This selected text information recognition component set will play an important role in the subsequent training process for the selected medical training text library, allowing the neural network to focus more on text recognition tasks related to the target medical information, avoiding interference from irrelevant components, thereby improving the accuracy and efficiency of training.

[0119] For example, in a large medical information integration project, there may be multiple medical training text sub-databases, such as those related to cardiovascular diseases, endocrine diseases, and nervous system diseases. For each sub-database, the medical information system needs to determine the corresponding set of selected text information recognition components in this way. If such a distinction is not made and all text information recognition components are involved in the training of each sub-database, it will cause confusion in the neural network during the training process. For example, when training the cardiovascular disease medical training text sub-database, those text information recognition components related to ophthalmic diseases may produce incorrect parameter adjustments due to receiving irrelevant text information, thereby affecting the entire neural network's ability to recognize medical information related to cardiovascular diseases.

[0120] As an implementation, the text information recognition component activation state indicator includes an activation mark and an inactivation mark, the activation mark and the inactivation mark are different, and the mark corresponding to the text information recognition component activation state indicator corresponds to the text information recognition component in the initial text processing neural network. Based on this, step S22, based on the text information recognition component activation state indicator, determines the selected text information recognition component set in each text information recognition component of the initial text processing neural network, which may include:

[0121] Step S221: determining the text information recognition component corresponding to the activation mark as the selected text information recognition component, and determining the selected text information recognition component to be in an activated state;

[0122] Step S222: the text information recognition component corresponding to the inactive mark is regarded as a non-selected text information recognition component, and the non-selected text information recognition component is determined to be in an inactive state; the non-selected text information recognition component is a text information recognition component other than the selected text information recognition component among the text information recognition components.

[0123] In step S221, in the system of regional medical information data integration and neural network training, the activation flag is an identifier in the text information recognition component activation status indicator, which plays a key role in determining which text information recognition components should be selected into the selected text information recognition component set for the selected medical training text sub-library.

[0124] For example, suppose there is an initial text processing neural network that contains multiple text information recognition components designed to recognize different types of medical information, such as disease symptoms, diagnosis results, treatment methods, etc. At the same time, there is a selected medical training text sub-library that focuses on medical information for a specific type of disease (such as respiratory diseases). The text in this sub-library contains content related to the symptoms (such as coughing, wheezing, etc.), diagnoses (such as the basis for the diagnosis of pneumonia and bronchitis), and treatments (such as the use of antibiotics, oxygen therapy, etc.) of patients with respiratory diseases.

[0125] The text information recognition component activation status indicator is determined based on the target medical information of the respiratory disease medical training text sub-library (i.e., information related to respiratory diseases) and the distribution of components in the initial text processing neural network (the original functional settings of each component). If the activation status indicator of a text information recognition component is an activation mark, it means that the component has some association with the target medical information in the respiratory disease medical training text sub-library and is suitable for participating in the training process of this sub-library.

[0126] Assume that there is a text information recognition component in the initial text processing neural network that is specifically used to recognize the symptoms of respiratory diseases. Since its function matches the target medical information (including symptom information) in the medical training text sub-library of respiratory diseases, its activation status indicator is marked as an activation flag. Based on this activation flag, the medical information system determines this text information recognition component that recognizes the symptoms of respiratory diseases as a selected text information recognition component and determines its status as an activated state. This process is targeted, ensuring that only components related to the target medical information can participate in subsequent operations on the medical training text sub-library.

[0127] Specifically, the medical information system can implement this determination process by querying a pre-established mapping relationship table. This mapping relationship table can store the association relationship between each text information recognition component and different medical training text sub-libraries, using an activation flag or other flag to indicate whether the component should be activated. For example, the rows in the table can represent different medical training text sub-libraries, the columns represent the various text information recognition components in the initial text processing neural network, and the values ​​in the cells are activation flags (such as 1) or inactivation flags (such as 0). When processing the medical training text sub-library for respiratory diseases, the medical information system traverses this table, finds the text information recognition component with a value of 1 in the corresponding column, determines it as the selected text information recognition component and sets it to the activated state.

[0128] In step S222, the activation mark corresponds to the inactivation mark. After determining the selected text information recognition component corresponding to the activation mark, the medical information system performs corresponding processing on the text information recognition components marked as inactivation marks.

[0129] Continuing with the above-mentioned respiratory disease medical training text sub-database as an example, there may be some text information recognition components in the initial text processing neural network that are not closely related to the target medical information of respiratory disease. For example, there is a component that is specifically used to recognize terms related to orthopedic diseases. Since its function is not related to the target medical information of the respiratory disease medical training text sub-database, its activation status indicator is marked as an inactive flag.

[0130] The medical information system regards this component for identifying orthopedic disease-related terms marked as inactive as a non-selected text information identification component and determines its status as inactive. This means that during the training process of the medical training text sub-library for respiratory diseases, this component will not participate in related operations, thus avoiding its interference with the training process.

[0131] From the perspective of technical implementation, it is also done by querying the mapping relationship table mentioned earlier. After the medical information system determines the selected text information recognition component (i.e., the component corresponding to the activation identifier), it determines the component with a value of 0 in the corresponding column in the table as a non-selected text information recognition component and marks them as inactive in the system. This processing method helps to improve the efficiency and accuracy of neural network training. If such a distinction is not made and all components are involved in the training, those components that are not related to the target medical training text database may make unnecessary parameter adjustments due to receiving irrelevant inputs, thereby affecting the performance of the entire neural network.

[0132] For example, in a large medical information integration project that includes multiple medical training text sub-databases (such as cardiovascular diseases, endocrine diseases, respiratory diseases, etc.), if the activated and inactivated text information recognition components are not accurately distinguished in this way for each sub-database, the neural network will be confused during the training process. Assuming that when training the respiratory disease medical training text sub-database, the text information recognition components related to other diseases (such as orthopedic diseases) are not set to an inactive state, then when the training text related to respiratory diseases is input, these unrelated components may incorrectly adjust their parameters and try to process the input text, even though they do not have the ability to process information related to respiratory diseases. This will not only waste computing resources, but may also cause the neural network to reduce the accuracy of recognizing medical information about respiratory diseases.

[0133] In summary, steps S221-S222 accurately screen out the selected text information recognition components related to the target medical information of the selected medical training text sub-library from the numerous text information recognition components of the initial text processing neural network through the identification and processing of the activation flag and the inactive flag, and determine the remaining components as inactive inactive text information recognition components. This process relies on technical means such as pre-established mapping relationship tables, which is crucial to improving the training effect and performance of neural networks in medical information data integration.

[0134] As an implementation manner, step S30, based on each text information recognition component and according to each first undetermined text segment corresponding to the selected medical training text, determines the first text information recognition error corresponding to the selected medical training text, which may include:

[0135] Step S31: obtaining the inferred medical information of each first to-be-determined text segment under each text information recognition component according to the to-be-determined text semantic vectors corresponding to each first to-be-determined text segment;

[0136] Step S32: for the first selected text information recognition component in the selected text information recognition component set corresponding to the selected medical training text sub-library, according to the inferred medical information of each first undetermined text segment under the first selected text information recognition component and the actual medical information corresponding to each first undetermined text segment, determine the text information recognition error of each first undetermined text segment under the first selected text information recognition component;

[0137] Step S33: taking the sum of the text information recognition errors of each first undetermined text segment under the first selected text information recognition component as the text information recognition error corresponding to the first selected text information recognition component;

[0138] Step S34: for the first non-selected text information recognition component in the non-selected text information recognition component set corresponding to the selected medical training text sub-library, determine the text information recognition error of each first pending text segment under the first non-selected text information recognition component according to the inferred medical information of each first pending text segment under the first non-selected text information recognition component and the second priori mark corresponding to each first pending text segment; wherein the second priori mark is used to indicate that the inferred medical information corresponding to the first pending text segment is the target medical information corresponding to the medical training text sub-library outside the non-selected medical training text sub-library;

[0139] Step S35: taking the sum of the text information recognition errors of each of the first undetermined text segments under the first non-selected text information recognition component as the text information recognition error corresponding to the first non-selected text information recognition component;

[0140] Step S36: Determine the first text information recognition error corresponding to the selected medical training text based on the text information recognition errors respectively corresponding to each selected text information recognition component in the selected text information recognition component set and the text information recognition errors respectively corresponding to each non-selected text information recognition component in the non-selected text information recognition component set.

[0141] In step S31, within the framework of regional medical information data integration and neural network training, the semantic vector of the pending text is a quantitative representation of the semantic information of the first pending text fragment. It contains rich semantic features and can help the medical information system determine the inferred medical information of each first pending text fragment under different text information recognition components.

[0142] For example, suppose the selected medical training text sub-database is about cardiovascular diseases, and the selected medical training text describes the condition of a patient, in which one of the first pending text segments is "the patient's electrocardiogram shows ST segment elevation, accompanied by chest pain symptoms". The semantic vector of the pending text corresponding to this segment is constructed through a series of technical means, and it may contain semantic information about symptoms (chest pain), examination results (electrocardiogram ST segment elevation) and the relationship between them.

[0143] Each text information recognition component in the initial text processing neural network has different functions. For example, some components are good at identifying the symptoms of cardiovascular disease, while others are good at inferring the type of disease based on the examination results. The medical information system uses the undetermined text semantic vectors corresponding to each first undetermined text segment and inputs them into each text information recognition component. For the text information recognition component that recognizes symptoms, based on the symptom-related semantic information in the undetermined text semantic vector, it can infer that the "chest pain" in this segment is symptom information related to cardiovascular disease; for the component that infers the type of disease based on the examination results, based on the examination result semantic information in the vector, it can infer that "electrocardiogram ST segment elevation" may indicate cardiovascular disease types such as myocardial infarction.

[0144] Specifically, the semantic vector of the pending text can be constructed by combining a word vector model (such as Word2Vec) with a semantic role labeling technology. First, the word vector model is used to convert the words in the text segment into vector representations, and then the semantic relationship between the words, such as the action performer and the action object, is determined through the semantic role labeling technology, thereby constructing a semantic vector of the pending text that can reflect the semantic structure of the text segment.

[0145] In step S32, the selected text information recognition component set is a specific component set corresponding to the selected medical training text sub-library, wherein the first selected text information recognition component plays an important role in identifying the first pending text segment related to the target medical information.

[0146] Continuing with the example of cardiovascular disease, assume that there is a first selected text information recognition component in the selected text information recognition component set that is specifically used to accurately determine the specific type of cardiovascular disease based on symptoms and examination results. For the first undetermined text segment mentioned above, "The patient's electrocardiogram showed ST segment elevation, accompanied by chest pain symptoms", the actual medical information is that the patient suffered from myocardial infarction.

[0147] Each first undetermined text segment has corresponding inference medical information under the first selected text information recognition component, such as this component infers that the patient may have myocardial infarction (hypothesis) based on the input undetermined text semantic vector. At this time, the medical information system can determine the text information recognition error by comparing the inference medical information with the actual medical information.

[0148] If a simple binary judgment method is used, when the inferred medical information is consistent with the actual medical information, the error is 0; when it is inconsistent, the error is 1. However, in actual applications, more complex measurement methods may be used. For example, cross-entropy can be used as an error measurement function. Let p(x) be the probability that the first selected text information recognition component predicts that the patient has myocardial infarction (assuming it is obtained through the Softmax function), and the actual medical information shows that the patient does have myocardial infarction, that is, the true probability is 1, then the cross-entropy formula is , where q(x)=1, so , this value is the text information recognition error of the first pending text segment under the first selected text information recognition component.

[0149] In step S33, after determining the text information recognition error of each first pending text segment under the first selected text information recognition component in the previous step, the medical information system sums these errors to obtain the text information recognition error corresponding to the first selected text information recognition component.

[0150] For example, suppose that in the selected medical training text, there are three first undetermined text segments T1, T2, and T3, and their text information recognition errors under the first selected text information recognition component are e1=0.2, e2=0, and e3=0.1 respectively (here it is assumed that a certain probability-based error measurement method is used). Then the text information recognition error corresponding to the first selected text information recognition component is E=e1+e2+e3=0.2+0+0.1=0.3.

[0151] This summation result comprehensively reflects the overall error of the first selected text information recognition component when processing all the first pending text segments. It takes into account the error contribution of each segment and provides a quantitative indicator for evaluating the performance of this component in identifying the target medical information in the selected medical training text.

[0152] In step S34, the non-selected text information recognition component set is a component set in the initial text processing neural network except the selected text information recognition component set. Although the first non-selected text information recognition component here is not directly related to the target medical information of the selected medical training text sub-library, its processing of the first pending text segment also needs to be considered when calculating the error.

[0153] Take a system containing two medical training text sub-databases of cardiovascular disease and endocrine disease as an example. For the selected medical training text sub-database of cardiovascular disease, it is assumed that there is a first non-selected text information recognition component that is specifically used to identify information related to endocrine disease. For the first undetermined text segment "the patient's electrocardiogram showed ST segment elevation, accompanied by chest pain symptoms" in the selected medical training text of cardiovascular disease, its second priori mark indicates that the information in this segment is not related to endocrine disease.

[0154] When the first non-selected text information recognition component processes the segment, it may output some inferred medical information. For example, it may mistakenly believe that some information in the segment is related to endocrine disease (hypothesis). The medical information system determines the text information recognition error based on the inferred medical information and the second priori marker.

[0155] If a binary judgment method similar to the previous one is used, when the inferred medical information is inconsistent with the situation indicated by the second priori marker (that is, it is inferred to be related to endocrine diseases, but in fact it is not related to endocrine diseases), the error is 1; when it is consistent, the error is 0. Other measurement methods can also be used, such as probability-based measurement. Assuming that the probability that this component outputs that this fragment is related to endocrine diseases is p, and according to the second priori marker, it is actually not related to endocrine diseases, that is, the true probability is 0, then the logarithmic loss function L = -log(1-p) can be used to calculate the text information recognition error of this fragment under the first non-selected text information recognition component.

[0156] Step S35 is similar to step S33. After calculating the text information recognition error of each first pending text segment under the first non-selected text information recognition component, the medical information system sums up these errors to obtain the text information recognition error corresponding to the first non-selected text information recognition component.

[0157] For example, for the three first undetermined text segments T1, T2, and T3 in the above cardiovascular disease selected medical training text, the text information recognition errors under the first non-selected text information recognition component (specially used to recognize endocrine disease related information) are e1=0.1, e2=0, and e3=0.2 respectively (assuming a probability-based error measurement method). Then the text information recognition error corresponding to the first non-selected text information recognition component is E=e1+e2+e3=0.1+0+0.2=0.3.

[0158] This result reflects the overall error of the first non-selected text information recognition component when processing the first pending text segment in the selected medical training text for cardiovascular diseases. Although this component is not directly related to the target medical information of cardiovascular diseases, this error indicator helps to comprehensively evaluate the performance of the neural network in processing the relationship between various components and different text segments.

[0159] In step S36, the medical information system determines the first text information recognition error corresponding to the entire selected medical training text based on the text information recognition errors corresponding to each component in the selected text information recognition component set and the text information recognition errors corresponding to each component in the non-selected text information recognition component set.

[0160] For example, suppose there are three components C1, C2, and C3 in the selected text information recognition component set, and their corresponding text information recognition errors are E1=0.3, E2=0.2, and E3=0.1 respectively; there are two components C4 and C5 in the non-selected text information recognition component set, and their corresponding text information recognition errors are E4=0.3 and E5=0.2 respectively.

[0161] The medical information system can use a weighted summation method to determine the first text information recognition error corresponding to the selected medical training text. Assume that the weight of the selected text information recognition component set is w1=0.6, and the weight of the non-selected text information recognition component set is w2=0.4 (the weight can be determined based on actual conditions, such as the importance of the components, etc.).

[0162] Then the first text information recognition error corresponding to the selected medical training text .

[0163] This first text information recognition error integrates the error conditions of the selected and non-selected text information recognition components when processing the first pending text segment in the selected medical training text, and comprehensively reflects the recognition error conditions of the neural network for the first pending text segment related to the target medical information when processing the selected medical training text, which provides an important basis for the subsequent adjustment of the neural network and helps to improve the accuracy and reliability of the neural network in the embodiment of the present invention.

[0164] In one embodiment, the method further includes a method for determining the first to-be-determined text segment and the second to-be-determined text segment, specifically including the following steps:

[0165] Step S1: acquiring an actual distribution position set corresponding to the selected medical training text according to the actual priori mark corresponding to the selected medical training text, wherein the actual distribution position set includes the actual distribution position of each target medical information corresponding to the selected medical training text;

[0166] Step S2: for a target pending text segment among multiple pending text segments corresponding to the selected medical training text, determine the position of the inferred medical information corresponding to the target pending text segment according to the pending text semantic vector corresponding to the target pending text segment; wherein the pending text segment is used to indicate a potential text segment of the target medical information corresponding to the medical training text in the medical training text;

[0167] Step S3: Calculate each of the inferred medical information positions corresponding to the target undetermined text segment and each actual distribution position in the actual distribution position set to obtain a matching coefficient between each of the inferred medical information positions corresponding to the target undetermined text segment and each actual distribution position; wherein the matching coefficient is used to indicate the fit between the inferred medical information position and the actual distribution position;

[0168] Step S4: if the maximum matching coefficient corresponding to the target undetermined text segment is greater than the first reference matching coefficient, the target undetermined text segment is used as the first undetermined text segment;

[0169] Step S5: If the maximum matching coefficient corresponding to the target undetermined text segment is smaller than the second reference matching coefficient, the target undetermined text segment is used as the second undetermined text segment; wherein the first reference matching coefficient is greater than the second reference matching coefficient.

[0170] In step S1, in the neural network training related to regional medical information data integration, the actual priori mark is a kind of identification information pre-set for the selected medical training text, which contains some prior knowledge about the target medical information in the text. The medical information system obtains the actual distribution position set based on this actual priori mark.

[0171] For example, suppose the selected medical training text is a medical record of a patient with cardiovascular disease. The actual prior mark may indicate the distribution pattern or specific identification of the target medical information in the entire medical record text, such as the disease diagnosis results (such as coronary heart disease), symptoms (such as chest pain, palpitations, etc.), and treatment methods (such as drug therapy, interventional therapy, etc.). By parsing this actual prior mark, the medical information system can determine the actual distribution location set of these target medical information in the medical record.

[0172] Specifically, the actual prior tags can be stored in a structured data format, such as XML or JSON format. It contains the mapping relationship between the target medical information type (such as disease name, symptoms, etc.) and the position information in the text (such as paragraph number, sentence number, etc.). The medical information system can use a parser (such as an XML parser or a JSON parser) to read the actual prior tags to obtain the actual distribution position set.

[0173] In step S2, the pending text segment is a portion that may contain target medical information separated from the selected medical training text. Each pending text segment has a corresponding pending text semantic vector, which is a quantitative representation of the segment semantics and contains rich semantic information.

[0174] Taking cardiovascular disease medical records as an example, there are multiple undetermined text segments in the selected medical training text. One of the target undetermined text segments may be the part describing the patient's symptoms, such as "the patient has often felt chest pain recently, and the pain radiates to the left shoulder." The undetermined text semantic vector corresponding to this segment contains semantic information about the symptoms (chest pain, radiation location, etc.).

[0175] The medical information system determines the location of the inferred medical information corresponding to the target text segment based on the semantic vector of the undetermined text. For example, by analyzing the semantic elements related to symptoms in the semantic vector and the relationship between them, the location of the inferred medical information "chest pain" in the segment (such as the starting position in the sentence, the index of the word, etc.) is determined.

[0176] The semantic vector of the pending text can be constructed by combining word vector technology (such as Word2Vec) in deep learning with semantic role labeling technology. First, the words in the fragment are converted into vector representations using word vector technology, and then the semantic relationship between words is determined through semantic role labeling to construct the semantic vector of the pending text.

[0177] In step S3, after obtaining the inferred medical information location and the actual distribution location set of the target undetermined text segment, the medical information system needs to calculate the matching coefficient between them. This matching coefficient can reflect the degree of fit between the inferred medical information location and the actual location where the target medical information should exist.

[0178] For example, for the symptom description segment in the cardiovascular disease medical record mentioned above, suppose the position of the inferred medical information "chest pain" in the segment is determined to start from the 5th word, and according to the actual distribution position set, the actual position of the symptom "chest pain" in the medical record should start from the 4th word. The medical information system can use a certain distance measurement method to calculate the matching coefficient.

[0179] A simple method is to use the absolute distance formula: matching coefficient = 1-|inferenced medical information location-actual distribution location| / maximum distance value. Assuming the maximum distance value is 10 (this value can be set based on factors such as the length of the text), then in this example, the matching coefficient = 1-|5-4| / 10=0.9.

[0180] In step S4, the first reference matching coefficient is a preset threshold value used to determine whether the target to-be-determined text segment is the first to-be-determined text segment.

[0181] Continuing with the example of the cardiovascular disease medical records above, if the calculated maximum matching coefficient is 0.9, and the first reference matching coefficient is set to 0.8, since 0.9>0.8, the target pending text segment is determined to be the first pending text segment. The first pending text segments are those pending text segments that have a high degree of fit with the actual distribution location of the target medical information, and they are likely to contain accurate target medical information.

[0182] If the maximum matching coefficient corresponding to the target undetermined text segment is smaller than the second reference matching coefficient, the target undetermined text segment is used as the second undetermined text segment, wherein the first reference matching coefficient is larger than the second reference matching coefficient.

[0183] The second reference matching coefficient is also a preset threshold value and is smaller than the first reference matching coefficient.

[0184] For example, if the maximum matching coefficient of the target undetermined text segment is 0.6, and the second reference matching coefficient is set to 0.7, since 0.6<0.7, the target undetermined text segment is determined to be the second undetermined text segment. The second undetermined text segments are those that have a low degree of fit with the actual distribution position of the target medical information and may contain information that is irrelevant to the target medical information or is erroneous.

[0185] Through this series of steps S1-S5, the medical information system can accurately distinguish the first pending text segment from the second pending text segment. This is of great significance for the subsequent error calculation based on different types of pending text segments (such as in steps S30 and S40) and the adjustment of the neural network. Accurately determining these two types of pending text segments helps to improve the recognition accuracy of the neural network for the target medical information, avoid the increase of training errors caused by incorrect treatment of different types of text segments, and thus improve the information processing effect in the whole regional medical information data integration process.

[0186] In actual medical information integration projects, there are a large number of medical training texts that need to be processed. For example, for a medical training text library covering multiple disease types (cardiovascular diseases, endocrine diseases, respiratory diseases, etc.), for each selected medical training text in each selected medical training text sub-library, it is necessary to accurately determine the first pending text segment and the second pending text segment according to such steps. Only in this way can we ensure that the neural network can correctly learn the characteristics of the target medical information during the training process and improve the ability to identify and integrate medical information.

[0187] At the same time, with the continuous updating of medical knowledge and the changes in the format of medical texts, the actual priori tags, the construction methods of the semantic vectors of the pending texts, and the reference matching coefficients may need to be adjusted. For example, with the emergence of new diagnostic criteria or symptom classifications for cardiovascular diseases, the identification and distribution information of the target medical information in the actual priori tags may need to be updated; if new terms or expressions are used in medical texts, the construction technology of the semantic vectors of the pending texts may need to be improved to ensure that the semantic information of the text fragments can be accurately represented; and the reference matching coefficients may need to be re-evaluated and set according to the new data distribution and the training effect of the neural network to achieve the best distinction between the first pending text fragment and the second pending text fragment, thereby improving the efficiency and accuracy of the entire medical information integration process.

[0188] In one embodiment, the method further includes other loss determination processes, which may specifically include:

[0189] Step S101: taking the actual distribution position corresponding to the first undetermined text segment under the maximum matching coefficient as the actual distribution position corresponding to the first undetermined text segment;

[0190] Step S102: determining the inferred medical information positions corresponding to the first undetermined text segments according to the undetermined text semantic vectors corresponding to the first undetermined text segments;

[0191] Step S103: Determine the position error corresponding to the selected medical training text based on the inferred medical information position and the actual distribution position corresponding to each first pending text segment, where the position error is used to indicate the loss between the inferred medical information position and the actual distribution position.

[0192] In step S101, in the neural network training system related to regional medical information data integration, the first undetermined text segment is a text segment with a high degree of fit with the target medical information selected in the previous step. For each first undetermined text segment, there is a maximum matching coefficient, which is obtained after matching calculation with the actual distribution position set, indicating the best matching degree between the segment and the actual position of the target medical information.

[0193] For example, suppose the selected medical training text is about the medical records of diabetic patients. One of the first pending text segments describes the patient's blood sugar index related information. In the previous matching calculation, this segment was matched with different positions in the actual distribution position set to obtain multiple matching coefficients, among which the actual distribution position corresponding to the maximum matching coefficient is a specific sentence position in a paragraph in the medical record. This position accurately points out the position where the blood sugar index information should exist in the medical record according to prior knowledge. The medical information system determines the actual distribution position under this maximum matching coefficient as the actual distribution position corresponding to the first pending text segment.

[0194] From the perspective of technical implementation, the medical information system can maintain a data structure to store the matching coefficient and the corresponding actual distribution position information of each first undetermined text segment. After calculating all the matching coefficients, the actual distribution position corresponding to the maximum matching coefficient is found by traversing this data structure, and it is marked as the actual distribution position of the first undetermined text segment.

[0195] In step S102, the semantic vector of the pending text is a quantitative representation of the semantic information of the first pending text segment, which contains rich semantic content and can reflect the semantic features related to medical information in the text segment. The medical information system determines the inferred medical information position corresponding to each first pending text segment based on this vector.

[0196] Take the first undetermined text segment in a medical record of a patient with hypertension as an example. This segment describes the patient's blood pressure measurement and some related symptoms. The semantic vector of the undetermined text is constructed by converting the words in the segment into vector representations and combining techniques such as semantic role labeling. This vector contains the semantic relationship between medical information such as blood pressure values ​​and symptoms.

[0197] The medical information system determines the position of the inferred medical information in the segment by analyzing the semantic vector of the pending text. For example, for the inferred medical information of blood pressure value, by analyzing the semantic elements related to blood pressure in the vector and their relationship with other elements, the specific position of the blood pressure value in the segment can be determined, such as which word in the sentence it starts.

[0198] To construct the semantic vector of the pending text, various techniques in natural language processing can be used. For example, a pre-trained word vector model (such as the GloVe model) is used to map the words in the text to a low-dimensional vector space, and then the semantic role labeling technology is used to determine the semantic role relationship between words, such as the relationship between the agent and the patient, so as to construct the semantic vector of the pending text that can accurately represent the semantics of the text.

[0199] In step S103, the medical information system calculates the position error corresponding to the selected medical training text by comparing the inferred medical information position of each first undetermined text segment with the actual distribution position determined in step S101. This position error reflects the degree of deviation between the inferred medical information position and the actual position, and is an important indicator for measuring the accuracy of the neural network in locating the target medical information position.

[0200] For example, suppose there are two first undetermined text segments in a medical record of a heart disease patient (selected medical training text). The inference position of the diagnosis name of heart disease (inferenced medical information) in the first segment starts at the 6th word of the sentence, while the actual distribution position determined according to step S101 starts at the 5th word; the inference position of the name of the therapeutic drug (inferenced medical information) in the second segment starts at the 10th word of the sentence, while the actual distribution position starts at the 8th word. There are many ways to calculate the position error. A simple method is to calculate the sum of the absolute differences, assuming that the position error is , for the i-th first undetermined text segment, the inferred medical information location is The actual distribution location is ,but , in this example, .

[0201] Another possible method is to use mean square error (MSE) to calculate the position error. The formula is , in the above example, n=2, then .

[0202] If the position error is large, it means that the neural network has a large deviation in identifying the location of the target medical information and needs to be adjusted. For example, in the process of processing a large amount of medical training text, if the overall position error remains high, it may mean that the structure or parameter setting of the neural network is not conducive to accurately capturing the location information of the target medical information. This may require adjusting the position prediction-related components in the neural network, or optimizing the training strategy of the entire neural network, such as adjusting the learning rate, increasing the number of training rounds, etc.

[0203] In actual medical information integration projects, many different types of medical training texts are involved, such as medical records and examination reports covering various types of diseases (endocrine diseases, respiratory diseases, nervous system diseases, etc.). For each selected medical training text, the position error is accurately calculated according to steps S101-S103. Moreover, with the continuous accumulation of medical data and the updating of medical knowledge, such as new disease diagnosis standards or changes in medical record formats, the construction of the semantic vector of the pending text, the determination of the actual distribution position, and the applicability of the position error calculation method may be affected. Therefore, the medical information system needs to continuously evaluate and adjust these links to ensure that the position error can accurately reflect the performance of the neural network in locating the target medical information position, thereby improving the accuracy and reliability of the entire regional medical information data integration process.

[0204] For example, as the understanding of certain rare diseases deepens, the information descriptions about these diseases in medical records may become more detailed and complex, which may require re-evaluating the construction method of the pending text semantic vector to better capture this new semantic information. At the same time, the new medical record format may change the actual distribution location of the target medical information, which requires updating the method for determining the actual distribution location, which in turn affects the calculation of the position error. Only in this way can we ensure that when the medical information system integrates medical information, the neural network can accurately identify and locate the target medical information, thereby improving the quality and efficiency of medical information integration.

[0205] In one embodiment, step S50, adjusting the initial text processing neural network according to the first text information recognition error and the second text information recognition error respectively corresponding to each medical training text in the medical training text library to obtain the target text processing neural network, may include:

[0206] Step S51: According to the position errors corresponding to each medical training text, the position prediction component in the initial text processing neural network is calibrated, according to the first text information recognition error and the second text information recognition error corresponding to each medical training text, the medical information prediction component in the initial text processing neural network is calibrated, and according to the position errors, the first text information recognition error and the second text information recognition error corresponding to each medical training text, the other network components in the initial text processing neural network are calibrated to obtain the target text processing neural network; wherein the medical information prediction component includes each text information recognition component of the initial text processing neural network.

[0207] The position prediction component is the part of the neural network used to predict the position of the target medical information in the text. It may have a hierarchical structure, such as an input layer receiving the feature representation of the pending text segment (such as the vector of the pending text semantic vector after dimensionality reduction or feature extraction), an intermediate layer containing multiple neurons for nonlinear transformation of features, and an output layer outputting the possible position information of the target medical information (such as the starting position index or relative position interval in the text).

[0208] For example, when processing a medical training text on cardiovascular disease, the position prediction component needs to predict the position of target medical information in the text, such as disease names (such as "coronary heart disease"), key symptoms (such as "chest pain"), etc.

[0209] The medical information system adjusts the position prediction component according to the position errors corresponding to each medical training text. When the position error is large, it means that the prediction result of the position prediction component deviates greatly from the actual target medical information position.

[0210] Assume that the gradient descent algorithm is used for calibration. The parameters in the position prediction component are (including parameters such as the weights of connecting neurons), the position error is E p (Such as the error obtained by calculating the inferred medical information location and the actual distribution location in the previous step). The basic formula of gradient descent is , where α is the learning rate, which determines the step size of each parameter update.

[0211] For example, if in a medical training text, the predicted position of the symptom "chest pain" deviates greatly from the actual position, resulting in a position error E p According to the above formula, the position prediction component will increase according to E p about The partial derivatives of (calculated by the back propagation algorithm) are used to adjust , thereby reducing the position error. As multiple medical training texts are processed, the position prediction component continuously adjusts its parameters and gradually improves the accuracy of the target medical information location prediction.

[0212] The medical information prediction component includes various text information recognition components of the initial text processing neural network, which are mainly used to recognize various target medical information contents in the medical training text, such as disease diagnosis, symptoms, treatment methods, etc. For example, one component may be specifically used to recognize the name of the disease, another component may be used to recognize the severity of symptoms, etc.

[0213] The medical information system adjusts the medical information prediction component according to the first text information recognition error and the second text information recognition error corresponding to each medical training text. The first text information recognition error reflects the recognition error when processing the first pending text segment related to the target medical information, and the second text information recognition error reflects the recognition error when processing the second pending text segment unrelated to the target medical information.

[0214] Assume that a similar gradient descent algorithm is used for calibration. For a text information recognition component in the medical information prediction component, its parameters are , if only the first text information recognition error is considered (To simplify the explanation, we will first look at the first text information recognition error. In actual calibration, both errors will be considered comprehensively.) The calibration formula is: .

[0215] For example, for a text information recognition component that recognizes cardiovascular disease symptoms, if when processing the first pending text segment containing the symptom description, "heart palpitations" are often mistakenly recognized as "heart panic" (causing the first text information recognition error to increase), then the above formula is used to adjust the parameter according to the partial derivative of the error with respect to the parameter , in order to improve the component's ability to correctly identify symptoms. When the first text information recognition error and the second text information recognition error are considered comprehensively, the two errors can be combined into a comprehensive error function (Where w1 and w2 are weights, set according to the actual situation to balance the impact of the two errors), then follow Make parameter adjustments.

[0216] In addition to the location prediction component and the medical information prediction component, other components in the initial text processing neural network may not directly perform location prediction or medical information content recognition, but they will also have an impact on the performance of the entire neural network. For example, there may be some components for data preprocessing (such as standardizing the input text), feature extraction (converting text features into a form suitable for neural network processing), or regulating network information flow (such as controlling the weight of information transmission between different layers).

[0217] The medical information system adjusts these other network components according to the position error, the first text information recognition error and the second text information recognition error corresponding to each medical training text, because the output of these components will indirectly affect the results of position prediction and medical information recognition.

[0218] Assume that the parameters of these other network components are , comprehensive error (Where w3, w4 and w5 are weights used to represent the different degrees of influence of position error, first text information recognition error and second text information recognition error on these components). The calibration formula is: .

[0219] For example, a component used for data preprocessing may affect subsequent location prediction and medical information recognition if it does not properly standardize text. Through the above tuning process, the parameters are adjusted according to the partial derivatives of the comprehensive error with respect to its parameters. , so that the performance of the entire neural network is optimized.

[0220] By adjusting the location prediction component, medical information prediction component and other network components respectively, the medical information system continuously optimizes the parameters of the initial text processing neural network. As the numerous medical training texts in the medical training text library are processed, the performance of the neural network gradually improves, and finally the target text processing neural network is obtained. The various text information recognition components in this target text processing neural network are configured to more accurately recognize the target medical information in the medical training text library, thereby improving the accuracy and efficiency of information processing in the process of regional medical information data integration.

[0221] Based on Figure 1 Based on the same principle as the method shown in , the present invention also provides a regional medical information data integration device 10, such as Figure 3 As shown, the device 10 comprises:

[0222] A text acquisition module 11 is used to acquire target medical information text;

[0223] The network calling module 12 is used to call the target text processing neural network, load the target medical information text into the target text processing neural network, recognize the target medical information text based on the target text processing neural network, and obtain the target medical information;

[0224] An information integration module 13, used to integrate the identified target medical information into a target space of a preset medical information library;

[0225] The network adjustment module 14 is used to adjust the target text processing neural network by the following process: obtaining a medical training text library, the medical training text library includes multiple medical training text sub-libraries, and the target medical information corresponding to each of the medical training text sub-libraries has errors; for a selected medical training text sub-library among the multiple medical training text sub-libraries, determining a selected text information recognition component set corresponding to the selected medical training text sub-library in each text information recognition component of the initial text processing neural network; wherein the selected text information recognition components in the selected text information recognition component set correspond to the target medical information corresponding to the selected medical training text sub-library; for a selected medical training text in the selected medical training text sub-library, based on the each text information recognition component, according to each first pending text segment corresponding to the selected medical training text, determining A first text information recognition error corresponding to a selected medical training text; wherein the first pending text segment is inferred to have target medical information corresponding to the medical training text; based only on the selected text information recognition components in the selected text information recognition component set, according to each second pending text segment corresponding to the selected medical training text, the second text information recognition error corresponding to the selected medical training text is determined; wherein the second pending text segment is inferred to not have target medical information corresponding to the medical training text; according to the first text information recognition error and the second text information recognition error respectively corresponding to each medical training text in the medical training text library, the initial text processing neural network is adjusted to obtain the target text processing neural network, and each text information recognition component in the target text processing neural network is configured to recognize the target medical information of the medical training text library.

[0226] The above embodiment introduces the regional medical information data integration device 10 from the perspective of a virtual module. The following describes a medical information system from the perspective of a physical module, as shown below:

[0227] The embodiment of the present invention provides a medical information system, such as Figure 4 As shown, the medical information system 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the medical information system 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the medical information system 100 does not constitute a limitation on the embodiments of the present invention.

[0228] An embodiment of the present invention provides a medical information system. The medical information system in the embodiment of the present invention includes: one or more processors; a memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above method is implemented.

Claims

1. A method for integrating regional medical information data, characterized in that: The method comprises: Obtain target medical information text; Retrieving a target text processing neural network, loading the target medical information text into the target text processing neural network, so as to recognize the target medical information text based on the target text processing neural network and obtain target medical information; Integrate the identified target medical information into the target space of the preset medical information library; The target text processing neural network is obtained by the following steps: Acquire a medical training text library, wherein the medical training text library includes a plurality of medical training text sub-libraries, and each of the medical training text sub-libraries respectively corresponds to target medical information with errors; For a selected medical training text sub-library among the multiple medical training text sub-libraries, determining a selected text information recognition component set corresponding to the selected medical training text sub-library in each text information recognition component of the initial text processing neural network; wherein the selected text information recognition components in the selected text information recognition component set correspond to the target medical information corresponding to the selected medical training text sub-library; For the selected medical training text in the selected medical training text sub-library, based on the respective text information recognition components, according to the respective first pending text segments corresponding to the selected medical training text, determine the first text information recognition error corresponding to the selected medical training text; wherein the first pending text segment is inferred to have target medical information corresponding to the medical training text; Based only on the selected text information recognition components in the selected text information recognition component set, according to each second pending text segment corresponding to the selected medical training text, the second text information recognition error corresponding to the selected medical training text is determined; wherein the second pending text segment is inferred to have no target medical information corresponding to the medical training text; According to the first text information recognition error and the second text information recognition error respectively corresponding to each medical training text in the medical training text library, the initial text processing neural network is adjusted to obtain the target text processing neural network, and each text information recognition component in the target text processing neural network is configured to recognize the target medical information of the medical training text library; Among the text information recognition components of the initial text processing neural network, determining the selected text information recognition component set corresponding to the selected medical training text sub-library includes: Obtaining an activation status indicator of a text information recognition component corresponding to the selected medical training text sub-library, wherein the text information recognition component activation status indicator is used to determine the activation status of each text information recognition component of the initial text processing neural network when calculating text information recognition errors; Determining the selected set of text information recognition components among the text information recognition components of the initial text processing neural network based on the text information recognition component activation state indicator; The text information recognition component activation status indicator is determined based on the target medical information corresponding to the medical training text library and the text information recognition component distribution corresponding to the initial text processing neural network.

2. The method according to claim 1, characterized in that The determining, based only on the selected text information recognition component in the selected text information recognition component set and according to each second pending text segment corresponding to the selected medical training text, the second text information recognition error corresponding to the selected medical training text comprises: Determine the priori marks corresponding to the respective second undetermined text segments as first priori marks, wherein the first priori marks are used to indicate that the inferred medical information corresponding to the second undetermined text segment is not the target medical information corresponding to the selected medical training text sub-library; For the first selected text information identification component in the selected text information identification component set, determine the inferred medical information of each of the second to-be-determined text segments under the first selected text information identification component according to the to-be-determined text semantic vectors corresponding to each of the second to-be-determined text segments; wherein the to-be-determined text semantic vectors are used to indicate the inferred medical information and the inferred medical information position corresponding to the to-be-determined text segment; Determine the text information recognition error of each of the second pending text segments under the first selected text information recognition component based on the inferred medical information of each of the second pending text segments under the first selected text information recognition component and the first priori label; The sum of the text information recognition errors of each of the second undetermined text segments under the first selected text information recognition component is used as the text information recognition error corresponding to the first selected text information recognition component; A second text information recognition error corresponding to the selected medical training text is determined based on the text information recognition errors respectively corresponding to the selected text information recognition components in the selected text information recognition component set.

3. The method according to claim 1, characterized in that: The text information recognition component activation state indicator includes an activation mark and a non-activation mark, the activation mark and the non-activation mark are different, and the mark corresponding to the text information recognition component activation state indicator corresponds to the text information recognition component in the initial text processing neural network; The step of determining the selected text information recognition component set from among the text information recognition components of the initial text processing neural network based on the text information recognition component activation state indicator comprises: Determine the text information recognition component corresponding to the activation mark as the selected text information recognition component, and determine the selected text information recognition component to be in an activated state; The text information recognition component corresponding to the inactive mark is used as a non-selected text information recognition component, and the non-selected text information recognition component is determined to be in an inactive state; the non-selected text information recognition component is a text information recognition component other than the selected text information recognition component among the various text information recognition components.

4. The method according to claim 1, characterized in that The determining, based on the respective text information recognition components and according to the respective first pending text segments corresponding to the selected medical training text, the first text information recognition error corresponding to the selected medical training text comprises: According to the undetermined text semantic vectors respectively corresponding to the first undetermined text segments, the inferred medical information of each of the first undetermined text segments under each of the text information recognition components is obtained; For the first selected text information recognition component in the selected text information recognition component set corresponding to the selected medical training text sub-library, determine the text information recognition error of each of the first pending text segments under the first selected text information recognition component based on the inferred medical information of each of the first pending text segments under the first selected text information recognition component and the actual medical information corresponding to each of the first pending text segments; The sum of the text information recognition errors of each of the first undetermined text segments under the first selected text information recognition component is used as the text information recognition error corresponding to the first selected text information recognition component; For the first non-selected text information recognition component in the non-selected text information recognition component set corresponding to the selected medical training text sub-library, determine the text information recognition error of each of the first pending text segments under the first non-selected text information recognition component based on the inferred medical information of each of the first pending text segments under the first non-selected text information recognition component and the second priori mark corresponding to each of the first pending text segments; wherein the second priori mark is used to indicate that the inferred medical information corresponding to the first pending text segment is not the target medical information corresponding to the medical training text sub-library other than the selected medical training text sub-library; The sum of the text information recognition errors of each of the first undetermined text segments under the first non-selected text information recognition component is used as the text information recognition error corresponding to the first non-selected text information recognition component; The first text information recognition error corresponding to the selected medical training text is determined based on the text information recognition errors respectively corresponding to each selected text information recognition component in the selected text information recognition component set and the text information recognition errors respectively corresponding to each non-selected text information recognition component in the non-selected text information recognition component set.

5. The method according to claim 1, characterized in that: The method further comprises: According to the actual priori mark corresponding to the selected medical training text, obtaining the actual distribution position set corresponding to the selected medical training text, wherein the actual distribution position set includes the actual distribution position of each target medical information corresponding to the selected medical training text; For a target pending text segment among the multiple pending text segments corresponding to the selected medical training text, determine the position of the inferred medical information corresponding to the target pending text segment according to the pending text semantic vector corresponding to the target pending text segment; wherein the pending text segment is used to indicate a potential text segment of the target medical information corresponding to the medical training text in the medical training text; Calculate each of the inferred medical information positions corresponding to the target undetermined text segment and each actual distribution position in the actual distribution position set to obtain a matching coefficient between each of the inferred medical information positions corresponding to the target undetermined text segment and each actual distribution position; wherein the matching coefficient is used to indicate the fit between the inferred medical information position and the actual distribution position; If the maximum matching coefficient corresponding to the target undetermined text segment is greater than the first reference matching coefficient, taking the target undetermined text segment as the first undetermined text segment; If the maximum matching coefficient corresponding to the target undetermined text segment is smaller than the second reference matching coefficient, taking the target undetermined text segment as the second undetermined text segment; Wherein, the first reference matching coefficient is greater than the second reference matching coefficient.

6. The method according to claim 5, characterized in that The method further comprises: Taking the actual distribution position corresponding to the first undetermined text segment under the maximum matching coefficient as the actual distribution position corresponding to the first undetermined text segment; Determine the inferred medical information positions corresponding to the first to-be-determined text segments respectively according to the to-be-determined text semantic vectors corresponding to the first to-be-determined text segments respectively; Based on the inferred medical information positions and actual distribution positions corresponding to each of the first pending text segments, the position error corresponding to the selected medical training text is determined, and the position error is used to indicate the loss between the inferred medical information position and the actual distribution position.

7. The method according to any one of claims 1 to 6, characterized in that: The adjusting the initial text processing neural network according to the first text information recognition error and the second text information recognition error respectively corresponding to each medical training text in the medical training text library to obtain the target text processing neural network includes: According to the position errors respectively corresponding to the various medical training texts, the position prediction component in the initial text processing neural network is calibrated, according to the first text information recognition error and the second text information recognition error respectively corresponding to the various medical training texts, the medical information prediction component in the initial text processing neural network is calibrated, and according to the position errors, the first text information recognition error and the second text information recognition error respectively corresponding to the various medical training texts, other network components in the initial text processing neural network are calibrated to obtain a target text processing neural network; Among them, the medical information prediction component includes various text information recognition components of the initial text processing neural network.

8. A regional medical information data integration device, characterized in that: The device comprises: A text acquisition module is used to acquire target medical information text; A network calling module, used to call a target text processing neural network, load the target medical information text into the target text processing neural network, recognize the target medical information text based on the target text processing neural network, and obtain target medical information; An information integration module, used to integrate the identified target medical information into a target space of a preset medical information database; The network calibration module is used to calibrate the target text processing neural network using the following process: Acquire a medical training text library, wherein the medical training text library includes a plurality of medical training text sub-libraries, and each of the medical training text sub-libraries respectively corresponds to target medical information with errors; For a selected medical training text sub-library among the multiple medical training text sub-libraries, a selected text information recognition component set corresponding to the selected medical training text sub-library is determined in each text information recognition component of the initial text processing neural network; wherein the selected text information recognition components in the selected text information recognition component set correspond to the target medical information corresponding to the selected medical training text sub-library; wherein, determining the selected text information recognition component set corresponding to the selected medical training text sub-library in each text information recognition component of the initial text processing neural network includes: obtaining an activation status indicator of the text information recognition component corresponding to the selected medical training text sub-library, wherein the text information recognition component activation status indicator is used to determine the activation status of each text information recognition component of the initial text processing neural network when calculating the text information recognition error; based on the text information recognition component activation status indicator, determining the selected text information recognition component set in each text information recognition component of the initial text processing neural network; wherein the text information recognition component activation status indicator is determined based on the target medical information corresponding to the medical training text library and the distribution of the text information recognition components corresponding to the initial text processing neural network; For the selected medical training text in the selected medical training text sub-library, based on the respective text information recognition components, according to the respective first pending text segments corresponding to the selected medical training text, determine the first text information recognition error corresponding to the selected medical training text; wherein the first pending text segment is inferred to have target medical information corresponding to the medical training text; Based only on the selected text information recognition components in the selected text information recognition component set, according to each second pending text segment corresponding to the selected medical training text, the second text information recognition error corresponding to the selected medical training text is determined; wherein the second pending text segment is inferred to have no target medical information corresponding to the medical training text; The initial text processing neural network is calibrated based on the first text information recognition error and the second text information recognition error respectively corresponding to each medical training text in the medical training text library to obtain the target text processing neural network, and each text information recognition component in the target text processing neural network is configured to recognize the target medical information of the medical training text library.

9. A medical information system, characterized in that: include: one or more processors; Memory; One or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 7 is implemented.

Images