Obstetrical midwifery decision-making model construction method based on mapping knowledge domain

Abstract

The invention relates to the technical field of model construction, in particular to an obstetric delivery decision model construction method based on a knowledge graph. The method comprises the following steps that multi-dimensional information collection is conducted on a pregnant woman to be delivered, individualized data are obtained, and the multi-dimensional information comprises a medical record text, a laboratory examination result, an ultrasonic image and fetal heart rate monitoring data; performing standardization and structural processing on the individualized data to generate a standardized multi-modal data set; performing entity identification and relation extraction processing on the standardized multi-modal data set according to a preset obstetrical domain ontology model to obtain an initial knowledge graph; and carrying out priori knowledge-based mapping expansion on the initial mapping knowledge domain to obtain a multi-modal mapping knowledge domain. Through an attention mechanism and a decision support interface, the interpretability of the model is enhanced, and doctors can understand and trust decision suggestions.

Description

Technical Field

[0001] The present invention relates to the technical field of model construction, and in particular to a method for constructing an obstetrics and midwifery decision-making model based on a knowledge graph. Background Art

[0002] Knowledge Graph is a technology that forms a structured network graph by organizing and representing a large amount of interrelated knowledge information. It can convert information into a machine-understandable form and show the relationship between entities (such as people, things, places, etc.). The core feature of the knowledge graph is to express knowledge in a graphical structure, where the nodes in the graph represent entities or concepts, and the edges represent the relationship between these entities or concepts. The main functions of the knowledge graph include: presenting complex information (such as text, database, etc.) in a structured form to make it easier to process and understand; by revealing the relationship between entities, it helps users establish connections between different concepts and form a richer context; the knowledge graph can help machines better understand the semantics of natural language, thereby improving the effectiveness of applications such as search, recommendation, and question answering. The obstetric midwifery decision model is an intelligent system based on clinical data and medical knowledge to help doctors make the best midwifery decisions in obstetrics. It provides support for the midwifery process by comprehensively considering multiple factors such as the health status of the pregnant woman, the condition of the fetus, the progress of labor, etc., helping doctors decide whether to choose natural delivery, midwifery surgery (such as cesarean section), or the use of auxiliary tools (such as forceps, vacuum extractor), as well as responding to emergencies during labor.

[0003] The traditional method of constructing obstetric midwifery decision-making models based on knowledge graphs often has the following problems: the situation of each pregnant woman is unique, and there may be multiple risk factors at the same time (such as advanced age for primiparity, gestational diabetes, multiple pregnancies, etc.); how to accurately represent these complex individual characteristics in the knowledge graph and design a model that can comprehensively evaluate multiple risk factors is a major challenge. This may require the combination of machine learning techniques to process high-dimensional feature spaces while maintaining the interpretability of the model. The medical data involved in obstetric midwifery decision-making includes not only text information, but also multimodal data such as image data (such as ultrasound images) and time series data (such as fetal heart monitoring data); integrating these different types of data into a unified knowledge graph for comprehensive midwifery decision support is a technically difficult area. Summary of the invention

[0004] Based on this, it is necessary for the present invention to provide a method for constructing an obstetric midwifery decision-making model based on a knowledge graph to solve at least one of the above-mentioned technical problems.

[0005] To achieve the above purpose, a method for constructing an obstetric midwifery decision-making model based on knowledge graph includes the following steps: Step S1: Collect multi-dimensional information from pregnant women to obtain individualized data, where the multi-dimensional information includes medical records, laboratory test results, ultrasound images, and fetal heart rate monitoring data; standardize and structure the individualized data to generate a standardized multimodal data set; Step S2: Perform entity recognition and relationship extraction on the standardized multimodal data set according to the preset obstetrics domain ontology model to obtain an initial knowledge graph; and perform graph expansion on the initial knowledge graph based on prior knowledge to obtain a multimodal knowledge graph; Step S3: constructing individual feature vectors of expectant pregnant women based on the multimodal knowledge graph, and performing low-dimensional vector space mapping based on graph embedding to obtain individual feature low-dimensional vectors; training a risk prediction model based on random forests based on the individual feature low-dimensional vectors, and adding an attention mechanism layer to obtain an individualized risk assessment model; Step S4: Design a text- and image-based data interface according to the obstetrics domain ontology model and the multimodal knowledge graph, and dynamically update the multimodal knowledge graph using the data interface to obtain a dynamically updated knowledge graph, wherein the data interface includes a data preprocessing module, a data fusion module, a graph update module, and a version control module; Step S5: Use the individualized risk assessment model to perform multi-version risk assessment on the dynamically updated knowledge graph, thereby obtaining risk assessment time series data; identify key risk factors through the attention mechanism layer of the individualized risk assessment model, thereby obtaining key risk factor data; Step S6: Perform key risk factor trend analysis on the risk assessment time series data based on the key risk factor data, and perform pan-mapping of the patient's individual characteristics to obtain risk factor-individual characteristic mapping data, thereby assisting doctors in making decisions.

[0006] The present invention can comprehensively and accurately obtain the health status of expectant mothers through multi-dimensional information collection, laying the foundation for subsequent analysis; standardization and structured processing enable data from different sources to be integrated in a unified format, thereby improving the availability and analysis efficiency of the data. The generated standardized multimodal data set provides a reliable input for the subsequent knowledge graph construction, improving the consistency and comparability of the data. Through entity recognition and relationship extraction, key medical entities (such as symptoms, medical records, test results, etc.) and their relationships can be extracted from the standardized multimodal data set to form a preliminary knowledge graph; the graph expansion based on prior knowledge further enriches the information content of the knowledge graph, enabling it to better reflect the actual clinical scenario, thereby supporting more complex knowledge reasoning and decision analysis. The construction of individual feature vectors enables the health information of expectant mothers to be expressed in vector form, which is convenient for calculation and analysis; low-dimensional vector space mapping simplifies the complexity of the data, making the subsequent risk prediction model more efficient. Using the random forest algorithm for risk prediction, combined with the attention mechanism, it is possible to more accurately identify the key factors affecting the risk, form an individualized risk assessment model, and help doctors make more accurate clinical decisions. The design of data interface enables multimodal knowledge graph to connect with external data sources in real time to maintain the timeliness and accuracy of information. Dynamic update of knowledge graph can reflect the latest clinical information and research results, and provide the latest data support for clinical decision-making; the design of each module ensures the standardization and consistency of data processing, and improves the maintainability and scalability of the system. Multi-version risk assessment can track the health changes of expectant mothers, identify potential risk factors in a timely manner, and provide a basis for clinical intervention. The attention mechanism can highlight key risk factors, so that doctors can pay attention to the most important health indicators when making decisions, thereby improving the effectiveness and accuracy of decisions. Through trend analysis, the law of risk factors changing over time can be identified to help doctors predict future health risks; the mapping relationship between risk factors and individual characteristics can enable doctors to better understand the unique situation of patients and formulate personalized intervention measures. This process not only improves the pertinence of medical services, but also strengthens patients' participation and understanding of their own health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings: Figure 1 It is a schematic diagram of the steps of the method for constructing an obstetrics and midwifery decision-making model based on a knowledge graph of the present invention; Figure 2 for Figure 1 Detailed step flow diagram of step S1; Figure 3 for Figure 1Detailed step flow chart of step S2 in FIG. DETAILED DESCRIPTION

[0008] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0009] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0010] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0011] To achieve this, please refer to Figures 1 to 3 The present invention provides a method for constructing an obstetric midwifery decision-making model based on a knowledge graph, the method comprising the following steps: Step S1: Collect multi-dimensional information from pregnant women to obtain individualized data, where the multi-dimensional information includes medical records, laboratory test results, ultrasound images, and fetal heart rate monitoring data; standardize and structure the individualized data to generate a standardized multimodal data set; Step S2: Perform entity recognition and relationship extraction on the standardized multimodal data set according to the preset obstetrics domain ontology model to obtain an initial knowledge graph; and perform graph expansion on the initial knowledge graph based on prior knowledge to obtain a multimodal knowledge graph; Step S3: constructing individual feature vectors of expectant pregnant women based on the multimodal knowledge graph, and performing low-dimensional vector space mapping based on graph embedding to obtain individual feature low-dimensional vectors; training a risk prediction model based on random forests based on the individual feature low-dimensional vectors, and adding an attention mechanism layer to obtain an individualized risk assessment model; Step S4: Design a text- and image-based data interface according to the obstetrics domain ontology model and the multimodal knowledge graph, and dynamically update the multimodal knowledge graph using the data interface to obtain a dynamically updated knowledge graph, wherein the data interface includes a data preprocessing module, a data fusion module, a graph update module, and a version control module; Step S5: Use the individualized risk assessment model to perform multi-version risk assessment on the dynamically updated knowledge graph, thereby obtaining risk assessment time series data; identify key risk factors through the attention mechanism layer of the individualized risk assessment model, thereby obtaining key risk factor data; Step S6: Perform key risk factor trend analysis on the risk assessment time series data based on the key risk factor data, and perform pan-mapping of the patient's individual characteristics to obtain risk factor-individual characteristic mapping data, thereby assisting doctors in making decisions.

[0012] In the embodiment of the present invention, reference Figure 1 The above is a schematic diagram of the steps of a method for constructing an obstetrics and midwifery decision-making model based on a knowledge graph of the present invention. In this example, the method for constructing an obstetrics and midwifery decision-making model based on a knowledge graph includes the following steps: Step S1: Collect multi-dimensional information from pregnant women to obtain individualized data, where the multi-dimensional information includes medical records, laboratory test results, ultrasound images, and fetal heart rate monitoring data; standardize and structure the individualized data to generate a standardized multimodal data set; The embodiment of the present invention obtains the medical record text data of pregnant women through a structured electronic medical record system, and the collected content covers basic information, pregnancy history, past medical history, etc. Then, with the help of the laboratory information system (LIS), the blood and urine test result data are extracted. In the third step, ultrasound image data, especially image information such as fetal structure ultrasound, placental position and amniotic fluid volume, are obtained through the integrated medical image archiving and communication system (PACS). Finally, fetal heart rate and fetal heart rate variability data are collected through fetal heart monitoring equipment. All these data are uniformly standardized and structured, including text segmentation, data normalization, denoising and other operations, and finally a standardized multimodal data set is generated. In the application scenario, important information in the medical record text, such as diagnosis, treatment process, drug use, etc., is extracted and formatted through natural language processing (NLP), laboratory data uses standard medical coding, such as LOINC or SNOMED, ​​and ultrasound images use image segmentation technology to extract specific areas for subsequent analysis.

[0013] Step S2: Perform entity recognition and relationship extraction on the standardized multimodal data set according to the preset obstetrics domain ontology model to obtain an initial knowledge graph; and perform graph expansion on the initial knowledge graph based on prior knowledge to obtain a multimodal knowledge graph; The embodiment of the present invention performs medical named entity recognition based on the ontology model in the obstetrics field. The text data is extracted through named entity recognition technology (NER) to extract entities such as symptoms, examination items, and treatment methods. Next, the relationship between entities is identified through a relationship extraction algorithm, such as the "examination-result" or "symptom-treatment" relationship, and entity relationship data is generated, and an initial knowledge graph is constructed with entities as nodes and relationships as edges. After that, based on the prior knowledge in the obstetrics field, the knowledge graph is expanded using a heterogeneous information network. The prior knowledge base involved includes known medical research, guidelines, and standards. The expansion process is implemented through an adaptive graph expansion algorithm to generate a multimodal knowledge graph. In specific applications, the expanded graph can cover more similar case information, such as extended data on pregnant women with similar medical histories and their related complications.

[0014] Step S3: constructing individual feature vectors of expectant pregnant women based on the multimodal knowledge graph, and performing low-dimensional vector space mapping based on graph embedding to obtain individual feature low-dimensional vectors; training a risk prediction model based on random forests based on the individual feature low-dimensional vectors, and adding an attention mechanism layer to obtain an individualized risk assessment model; The embodiment of the present invention extracts individual characteristics of pregnant women and generates individual feature vectors of expectant pregnant women. The information involved includes basic physiological parameters of pregnant women, historical pregnancy and childbirth experience, etc. Then, these high-dimensional information are mapped to low-dimensional vector space using graph embedding technology, and the generation of low-dimensional vectors is performed through feature extraction and embedding by graph neural network (GNN). The low-dimensional vector is used as input to train a risk prediction model based on the random forest algorithm, where the random forest consists of multiple decision trees, and each tree independently predicts the potential risk factors of pregnant women. In order to improve the performance of the model, a graph attention mechanism layer is added to dynamically adjust the importance of the features, and finally generate an individualized risk assessment model by weighting the importance of the features. In actual scenarios, the dimension of the low-dimensional vector can be set to 128, the number of trees in the random forest is set to 100, and the attention mechanism is optimized based on the correlation of historical complications of pregnant women.

[0015] Step S4: Design a text- and image-based data interface according to the obstetrics domain ontology model and the multimodal knowledge graph, and dynamically update the multimodal knowledge graph using the data interface to obtain a dynamically updated knowledge graph, wherein the data interface includes a data preprocessing module, a data fusion module, a graph update module, and a version control module; The embodiment of the present invention designs a data interface based on text and images to realize the dynamic update of the multimodal knowledge graph. First, the text data is segmented and denoised by natural language processing (NLP) technology, and the ultrasound image is denoised and edge detected by convolutional neural network (CNN). Then, the new data is standardized and cleaned through the preprocessing module of the data interface, the new multimodal data is integrated with the existing knowledge graph, and the graph update module is used to add new entities, update relationships and modify attributes. After each knowledge graph is updated, the version control module records the updated version information to ensure that the system can track the evolution path of the graph. In the application scenario, the preprocessing module will process the newly generated fetal heart monitoring data in real time, and a new version label will be generated after each graph update for subsequent version backtracking and analysis.

[0016] Step S5: Use the individualized risk assessment model to perform multi-version risk assessment on the dynamically updated knowledge graph, thereby obtaining risk assessment time series data; identify key risk factors through the attention mechanism layer of the individualized risk assessment model, thereby obtaining key risk factor data; The embodiment of the present invention uses an individualized risk assessment model to perform multi-version risk assessment on the updated knowledge graph, and the system will dynamically adjust the risk assessment based on the latest data of the pregnant woman. Time series data for risk assessment is generated through multiple time points, and the risk change trend at different time points is analyzed using time series analysis methods. The attention mechanism layer identifies the importance of each risk factor and outputs key risk factor data, such as high-risk factors that may lead to complications. In the application scenario, the system will update the risk assessment once a day during the pregnant woman's hospitalization, and the output key risk factor data includes features such as the pregnant woman's blood pressure and fetal heart rate abnormalities.

[0017] Step S6: Perform key risk factor trend analysis on the risk assessment time series data based on the key risk factor data, and perform pan-mapping of the patient's individual characteristics to obtain risk factor-individual characteristic mapping data, thereby assisting doctors in making decisions.

[0018] The embodiment of the present invention performs trend analysis on time series data based on key risk factor data. The system first aligns the spatiotemporal information of risk changes, captures the change pattern of risk factors, and maps different individual characteristics to relevant risk factors through pan-mapping technology. The generated risk factor-individual characteristic mapping data is used to help doctors make comprehensive judgments and decisions on risk factors. In actual applications, the system will perform multi-scale analysis on long-term monitoring data and generate risk curves for different time periods. For example, excessive fetal weight in late pregnancy may cause the risk of dystocia, thereby helping doctors take corresponding preventive measures in advance.

[0019] The present invention can comprehensively and accurately obtain the health status of expectant mothers through multi-dimensional information collection, laying the foundation for subsequent analysis; standardization and structured processing enable data from different sources to be integrated in a unified format, thereby improving the availability and analysis efficiency of the data. The generated standardized multimodal data set provides a reliable input for the subsequent knowledge graph construction, improving the consistency and comparability of the data. Through entity recognition and relationship extraction, key medical entities (such as symptoms, medical records, test results, etc.) and their relationships can be extracted from the standardized multimodal data set to form a preliminary knowledge graph; the graph expansion based on prior knowledge further enriches the information content of the knowledge graph, enabling it to better reflect the actual clinical scenario, thereby supporting more complex knowledge reasoning and decision analysis. The construction of individual feature vectors enables the health information of expectant mothers to be expressed in vector form, which is convenient for calculation and analysis; low-dimensional vector space mapping simplifies the complexity of the data, making the subsequent risk prediction model more efficient. Using the random forest algorithm for risk prediction, combined with the attention mechanism, it is possible to more accurately identify the key factors affecting the risk, form an individualized risk assessment model, and help doctors make more accurate clinical decisions. The design of data interface enables multimodal knowledge graph to connect with external data sources in real time to maintain the timeliness and accuracy of information. Dynamic update of knowledge graph can reflect the latest clinical information and research results, and provide the latest data support for clinical decision-making; the design of each module ensures the standardization and consistency of data processing, and improves the maintainability and scalability of the system. Multi-version risk assessment can track the health changes of expectant mothers, identify potential risk factors in a timely manner, and provide a basis for clinical intervention. The attention mechanism can highlight key risk factors, so that doctors can pay attention to the most important health indicators when making decisions, thereby improving the effectiveness and accuracy of decisions. Through trend analysis, the law of risk factors changing over time can be identified to help doctors predict future health risks; the mapping relationship between risk factors and individual characteristics can enable doctors to better understand the unique situation of patients and formulate personalized intervention measures. This process not only improves the pertinence of medical services, but also strengthens patients' participation and understanding of their own health management.

[0020] Preferably, step S1 comprises the following steps: Step S11: collecting basic information, pregnancy history, and past medical history information from pregnant women through a unified data collection form, thereby obtaining medical record text data; Step S12: Obtaining blood and urine test results through a laboratory information system to obtain laboratory test result data; Step S13: using an integrated medical image archiving and communication system to obtain ultrasound image data, including ultrasound of fetal structure, placental position, and amniotic fluid volume; Step S14: collecting fetal heart rate and fetal heart rate variability information from the expectant pregnant woman through a fetal heart rate monitoring device, thereby obtaining fetal heart rate monitoring data; Step S15: cleaning the medical record text data, laboratory test result data, ultrasound image data, and fetal heart rate monitoring data, and merging them into individualized data; Step S16: Standardize and structure the individualized data to generate a standardized multimodal data set.

[0021] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed step flow diagram of step S1 in the embodiment of the present invention, step S1 includes the following steps: Step S11: collecting basic information, pregnancy history, and past medical history information from pregnant women through a unified data collection form, thereby obtaining medical record text data; The embodiment of the present invention collects the basic information, pregnancy history and medical history of expectant mothers by designing a unified data collection form. Basic information includes name, age, height, weight, blood type, etc.; pregnancy history information involves whether there has been miscarriage, premature birth, cesarean section, etc.; past medical history covers chronic diseases such as diabetes, hypertension, thyroid disease, etc. When the nurse or doctor fills out the form during the first prenatal check-up of the pregnant woman, the medical record text data will be automatically stored in the electronic medical record system (EMR), and the system ensures the accuracy and consistency of the data through preset fields. For example, the age field is set to integer type, and the medical history field uses multiple selections or drop-down boxes to avoid errors caused by manual input. In the application scenario, during the first visit of the pregnant woman, the doctor fills in the relevant data through the interactive system to generate an electronic medical record text.

[0022] Step S12: Obtaining blood and urine test results through a laboratory information system to obtain laboratory test result data; The embodiment of the present invention uses a laboratory information system (LIS) to automatically extract blood and urine test result data, including blood routine, liver function, kidney function, blood sugar and other blood indicators, as well as proteinuria, glucosuria and other urine indicators. After the laboratory data is generated, the LIS system directly pushes the data to the central database through a standardized interface without manual entry. In the laboratory results, each indicator will be accompanied by a unit and a normal reference range, and the system will automatically compare and mark abnormal data. In actual operation, the blood and urine test results of pregnant women during each prenatal checkup will be uploaded in real time, and a report will be generated within a few hours, which the doctor can view and confirm through the LIS system.

[0023] Step S13: using an integrated medical image archiving and communication system to obtain ultrasound image data, including ultrasound of fetal structure, placental position, and amniotic fluid volume; In the embodiment of the present invention, during the ultrasound examination, the ultrasound equipment automatically stores the ultrasound image data to the hospital's central server through the integrated medical image archiving and communication system (PACS). The image data includes key information such as the fetal structural ultrasound, placental position, and amniotic fluid volume. By setting a specific image acquisition protocol, the system ensures the clarity and uniform format of the image, which is convenient for subsequent processing and analysis. The ultrasound physician will evaluate the growth and development of the fetus, the position of the placenta, and the amniotic fluid according to the standard operating procedures, and record the evaluation results in the system. Each ultrasound image is automatically associated with the patient's file and can be integrated with the medical record text and laboratory test results. In actual scenarios, the ultrasound image of the pregnant woman will be uploaded after each prenatal examination, and will be further evaluated and annotated by the doctor.

[0024] Step S14: collecting fetal heart rate and fetal heart rate variability information from the expectant pregnant woman through a fetal heart rate monitoring device, thereby obtaining fetal heart rate monitoring data; The embodiment of the present invention collects data of fetal heart rate (FHR) and fetal heart rate variability (FHRvariability) through a fetal heart monitoring device. During the monitoring process, the device monitors the heartbeat of the fetus through an abdominal patch and records the fetal heart rate curve in real time. The system analyzes the characteristics of fetal heart rate variability, acceleration and deceleration through a built-in algorithm to generate a monitoring report. After the information collection is completed, the fetal heart monitoring device automatically uploads the monitoring data to the central database, and the data is stored as structured time series data, including instantaneous fluctuations and average values ​​of the fetal heart rate. In actual scenarios, pregnant women undergo fetal heart monitoring every day when they are hospitalized in the obstetrics department, and the monitoring time is set to 30 minutes to 1 hour to ensure continuous monitoring of the fetal status.

[0025] Step S15: cleaning the medical record text data, laboratory test result data, ultrasound image data, and fetal heart rate monitoring data, and merging them into individualized data; After all data collection is completed, the embodiment of the present invention first cleans the medical record text, laboratory test results, ultrasound images and fetal heart monitoring data. Text data is cleaned by natural language processing (NLP) technology to remove redundant information and typos, and the same data is merged; laboratory test result data is processed by screening invalid values ​​and outliers; ultrasound image data is noise-removed by edge detection algorithm; fetal heart monitoring data is interpolated to complete the missing data in the monitoring process. All these data will be merged according to individual ID after cleaning to generate a complete individualized data set. In the application scenario, the data cleaning process is automated, and the system will detect invalid or outliers and issue an alarm to remind the doctor to check.

[0026] Step S16: Standardize and structure the individualized data to generate a standardized multimodal data set.

[0027] After the individualized data is cleaned, the embodiment of the present invention will be standardized and structured. Text data is standardized through word segmentation and named entity recognition (NER) technology, laboratory test data is uniformly encoded according to standard medical coding systems (such as LOINC or SNOMED CT), ultrasound image data is extracted and classified through convolutional neural networks (CNN), and fetal heart monitoring data is standardized at fixed time intervals (such as taking the average fetal heart rate every 5 seconds). Structured data will be stored in a database to form a standardized multimodal data set. In the application scenario, the data processing process will be automatically adjusted according to the specific status of the pregnant woman and the frequency of examination to ensure that the standardized data set can be dynamically updated and kept up to date.

[0028] The present invention can standardize the input of information and reduce errors in manual input through a unified data collection form. This method ensures the comprehensiveness and consistency of information, and is conducive to the standardization and analysis of subsequent data processing. In addition, the medical record text data can cover a wide range of health information, including past medical history and changes during pregnancy, providing a basis for obstetric decision-making. Laboratory test results usually include blood and urine analysis, which can reveal the internal health status of pregnant women, such as whether there is anemia, infection or other metabolic abnormalities. These data are crucial for risk assessment, especially in the field of obstetrics, which can detect potential risk factors for pregnant women and fetuses early; the accuracy and timeliness of data are ensured through the automated data acquisition of the laboratory information system. Ultrasound image data is an important indicator for evaluating fetal development, placental position and amniotic fluid volume. Through the integrated picture archiving and communication system (PACS), image data can be effectively managed and retrieved to ensure the integrity and traceability of the data. At the same time, through this system, the phenomenon of information islands can be avoided, and the sharing of image data between different departments can be ensured, which is convenient for doctors to fully understand the conditions of pregnant women and fetuses. Fetal heart rate and variability are important indicators for assessing fetal health. Automated data collection through fetal heart monitoring equipment can monitor fetal heart activity in real time and continuously, and detect dangerous signals such as fetal distress early. This automated equipment reduces the subjective errors of manual monitoring and provides more objective and accurate fetal health data. Data cleaning is a key step to ensure data quality, which can effectively remove duplicate, erroneous or missing information, thereby improving data accuracy. By cleaning and merging multimodal data to form an individualized data set, it is ensured that the comprehensive health information of expectant mothers is displayed in an integrated data set, providing a high-quality data foundation for subsequent decision model training. Standardization processing ensures consistency between different data sources (such as text, laboratory results, images, and monitoring data), so that data can be analyzed and processed uniformly. Structured processing helps to convert complex multimodal data into a processable form, which is convenient for subsequent machine learning and data analysis operations. The generated standardized multimodal data set provides high-quality and unified input for the construction of the decision model, improving the accuracy and reliability of the model.

[0029] Preferably, step S2 comprises the following steps: Step S21: performing medical named entity recognition on the text in the standardized multimodal dataset according to a preset obstetrics domain ontology model, thereby obtaining entity recognition result data; Step S22: extracting the relationship between medical entities from the entity recognition result data using the preset obstetrics domain ontology model, and determining the semantic relationship between entities, thereby obtaining entity relationship data; Step S23: Structuring entities and relationships according to entity recognition result data and entity relationship data, thereby constructing an initial knowledge graph, in which entities serve as nodes of the graph and relationships serve as edges of the graph; Step S24: Based on the prior knowledge in the obstetrics field, the initial knowledge graph is adaptively expanded based on a heterogeneous information network, thereby obtaining a knowledge expansion graph; Step S25: mapping the association between non-text information and graph nodes of the standardized multimodal data set according to the knowledge extension graph, thereby obtaining a multi-source data fusion graph; Step S26: Perform consistency check on the multi-source data fusion graph and perform conflict resolution to obtain a multimodal knowledge graph.

[0030] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 1 Detailed step flow diagram of step S2 in the embodiment of the present invention, step S2 includes the following steps: Step S21: performing medical named entity recognition on the text in the standardized multimodal dataset according to a preset obstetrics domain ontology model, thereby obtaining entity recognition result data; The embodiment of the present invention processes part of the text data in the standardized multimodal data set. A preset ontology model based on the obstetrics field is used, which contains professional terms and definitions related to obstetrics, and the key information in the text, such as disease names, drug names, examination items, symptom descriptions, etc., is identified through medical named entity recognition (NER) technology. In specific operations, a deep learning model (such as BERT or BioBERT) is used to semantically parse the text, and the model will automatically annotate it in combination with the entity labels in the obstetrics ontology. In the application scenario, after the text data of the medical records of pregnant women is processed, the system can automatically identify medical entities such as "gestational diabetes" and "placenta previa", and finally generate entity recognition result data.

[0031] Step S22: extracting the relationship between medical entities from the entity recognition result data using the preset obstetrics domain ontology model, and determining the semantic relationship between entities, thereby obtaining entity relationship data; After completing entity recognition, the embodiment of the present invention uses the obstetrics domain ontology model to further analyze the semantic relationship between each entity. The interactive relationship between entities, such as "disease-symptoms", "drug-indications", "examination-results", etc., is identified from the text through a medical entity relationship extraction algorithm (such as based on a dependency tree or a graph convolutional network GCN), and its relationship type is determined. During operation, the system automatically analyzes and annotates the relationship between entities according to preset rules (such as based on logical reasoning or statistical association). In the application scenario, the system will identify entity relationships such as "there is a correlation between gestational hypertension and fetal growth restriction" and generate structured entity relationship data.

[0032] Step S23: Structuring entities and relationships according to entity recognition result data and entity relationship data, thereby constructing an initial knowledge graph, in which entities serve as nodes of the graph and relationships serve as edges of the graph; The embodiment of the present invention constructs an initial knowledge graph based on the identified entities and relationships. Entities are used as nodes, and relationships are used as edges in the graph. The identified obstetric entities and relationships are visualized and structured for storage using a graph data structure. During operation, a graph database (such as Neo4j or Stardog) is used to store and manage the recognition results. Each entity node includes corresponding attributes. For example, a disease entity node contains the name of the disease, the probability of disease, etc., and the relationship edge is marked with the corresponding relationship type and directionality. In actual scenarios, the initial knowledge graph can display relationships such as "gestational diabetes-fetal weight", which makes it easier for doctors to understand the relationship between complex medical information.

[0033] Step S24: Based on the prior knowledge in the obstetrics field, the initial knowledge graph is adaptively expanded based on a heterogeneous information network, thereby obtaining a knowledge expansion graph; In order to further expand the expressive power of the graph, the embodiment of the present invention uses an adaptive graph expansion method based on a heterogeneous information network to introduce prior knowledge in the obstetric field into the initial knowledge graph. The specific method is to integrate relevant information from other medical fields (such as cardiovascular diseases, metabolic diseases, etc.) into the graph through a deep learning model or logical reasoning rules, and perform adaptive expansion of nodes and edges. The system will automatically infer potential associations based on existing entities and relationships, such as the potential connection between hyperglycemia and pregnancy complications. In the application scenario, the system will infer some unlabeled risk relationships through existing knowledge, such as "gestational diabetes may increase the risk of postpartum hemorrhage", thereby obtaining a knowledge expansion graph.

[0034] Step S25: mapping the association between non-text information and graph nodes of the standardized multimodal data set according to the knowledge extension graph, thereby obtaining a multi-source data fusion graph; Based on the knowledge expansion graph, the embodiment of the present invention further associates and maps non-text information (such as ultrasound images, fetal heart monitoring data) with knowledge graph nodes for standardized multimodal data sets. The specific operation includes using image analysis algorithms (such as CNN) to extract fetal growth and development features from ultrasound images and mapping them to corresponding graph nodes. For time series data (such as fetal heart monitoring data), time series analysis tools (such as LSTM) are used to associate key indicators with related entities in the knowledge graph (such as "fetal health status"). In the application scenario, the system can automatically map the placenta position information in the ultrasound image to the "placenta previa" node in the knowledge graph to form a multi-source data fusion graph.

[0035] Step S26: Perform consistency check on the multi-source data fusion graph and perform conflict resolution to obtain a multimodal knowledge graph.

[0036] After the multi-source data fusion graph is generated in the embodiment of the present invention, the system will perform consistency checks and conflict resolution. A graph verification algorithm (such as graph structure verification or rule-based conflict detection algorithm) is used to detect whether there are duplications, contradictions or logical errors in the nodes and edges in the graph. For possible conflicts, such as inconsistent attribute values ​​of a certain entity node in different data sources, the system will automatically resolve conflicts through preset priority rules (such as based on the credibility of the data source). In the application scenario, the system will detect the inconsistency between fetal heart monitoring data and ultrasound examination results in fetal health assessment, and resolve the conflict through doctor confirmation, and finally generate a multimodal knowledge graph with high consistency.

[0037] The medical named entity recognition (NER) of the present invention can automatically extract clinically relevant medical entities (such as diseases, symptoms, drugs, etc.) from the text, thereby making the data more structured and operable. This step not only improves the efficiency of data processing, but also ensures the accuracy of information, laying a solid foundation for subsequent relationship extraction and knowledge graph construction. Relationship extraction is a key link in building a knowledge graph. By identifying the semantic relationships between medical entities (such as causal relationships, similarities, etc.), the connection between entities can be revealed. This provides the necessary edges for the formation of a knowledge graph, so that the knowledge graph not only contains node information, but also has rich structured relationships, which is helpful for subsequent data analysis and reasoning. Through structured processing, entities and relationships are effectively organized into a knowledge graph, and this graphical representation is intuitive and easy to understand. The initial knowledge graph not only helps to quickly access and query relevant medical knowledge, but also provides a basic structure for subsequent data analysis and intelligent reasoning, and enhances the overall information retrieval capability of the system. By combining prior knowledge in the field of obstetrics, the initial knowledge graph can be adaptively expanded to make it more comprehensive and accurate. Heterogeneous information networks can integrate multiple types of data sources and provide a broader perspective, making knowledge graphs more reliable in clinical decision support and improving their applicability and flexibility. Through association mapping, non-text information (such as images, experimental results, etc.) in standardized multimodal datasets can be combined with nodes in the knowledge graph to form a multi-source data fusion graph. This process enhances the interconnection between different data sources, allowing multimodal information to be comprehensively utilized in clinical applications, providing more comprehensive decision support for medical staff. Consistency checking and conflict resolution are key steps to ensure the quality of knowledge graphs. By resolving contradictions between different data sources, the reliability and accuracy of multimodal knowledge graphs can be improved. This high-quality knowledge graph provides a more solid foundation for clinical analysis, decision support, and training of machine learning models, helping medical staff make more scientific and effective clinical decisions.

[0038] Preferably, step S24 includes the following steps: Step S241: performing heterogeneous information network modeling on the initial knowledge graph to obtain a heterogeneous information network model, wherein the heterogeneous information network modeling specifically involves constructing a heterogeneous graph structure including multiple node types and edge types; The embodiment of the present invention performs heterogeneous information network modeling on the initial knowledge graph. The core of heterogeneous information network modeling is to construct a graph structure containing multiple node types and edge types, where nodes can represent different entity types in the obstetric field (such as diseases, examinations, drugs, etc.), and edges represent multiple semantic relationships between these entities (such as "drugs-disease treatment", "examinations-test results", etc.). In specific operations, a graph database (such as Neo4j) is used to store and organize these heterogeneous information, and ensure that each node and edge can reflect its unique attributes and relationships. In practical applications, the constructed heterogeneous graph structure can include nodes for examination indicators related to fetal health and nodes such as fetal growth and development, and are connected through edges such as examination results and medical history information.

[0039] Step S242: designing a multi-view attention mechanism based on prior knowledge in the obstetrics field, and assigning weights to nodes and edges in the heterogeneous information network model, thereby obtaining a weighted heterogeneous information network; After constructing a heterogeneous information network model, the embodiment of the present invention designs a multi-view attention mechanism based on prior knowledge in the obstetric field. The multi-view attention mechanism assigns different weights to nodes and edges through different perspectives (such as medical history, examination results, ultrasound images, etc.), thereby highlighting important information in the graph. In operation, a multi-view attention model based on a graph convolutional network (GCN) is used to integrate information from different data sources into the graph network, and different weights are assigned to each node according to its influence in the graph. When setting specific parameters, weights can be assigned according to the accuracy and reliability of different data sources. For example, fetal heart monitoring data may have a higher weight. In the application scenario, the system will focus on nodes and relationships that may affect obstetric risk assessment based on the combination of the medical history of the pregnant woman and the results of fetal heart monitoring, and obtain a weighted heterogeneous information network.

[0040] Step S243: performing representation learning based on a graph neural network on the weighted heterogeneous information network, and generating low-dimensional embedding vectors of the nodes, thereby obtaining a network embedding representation; The embodiment of the present invention performs representation learning based on a graph neural network (GNN) on a weighted heterogeneous information network to generate a low-dimensional embedding vector of the node. In this process, a graph neural network model based on a heterogeneous network (such as HeteroGNN) is used to aggregate information for each node and its neighboring nodes, and a low-dimensional embedding vector is generated through a convolution operation. In the specific steps, the weighted network is firstly subjected to feature extraction, and the extracted features are processed through several layers of neural networks to finally generate a low-dimensional embedding vector with semantic meaning. In actual scenarios, the embedding vector can represent the potential characteristics of obstetric risk factors, such as certain examination indicators of pregnant women may imply the risk of high-risk pregnancy.

[0041] Step S244: inferring and predicting entity relationships in the initial knowledge graph according to the network embedding representation, thereby obtaining candidate extended knowledge data; The embodiment of the present invention infers and predicts the entity relationships in the initial knowledge graph based on the network embedding representation. In this step, the system uses the learning results of the graph neural network to infer the entities and relationships in the graph, and predicts the relationships that may exist but are not explicitly labeled. In specific operations, a graph reasoning algorithm (such as a path-based reasoning algorithm or a message passing algorithm) is used to analyze the low-dimensional embedding vector to infer new entity relationships. For example, the possible obstetric complications of a pregnant woman can be inferred from her examination results and medical history. In actual applications, the system may infer risk relationships that are not explicitly stated in the original data, such as "pregnancy-induced hypertension may lead to premature birth", and generate candidate extended knowledge data.

[0042] Step S245: Using a multi-task learning framework, the candidate extended knowledge data is screened and verified based on the simultaneous optimization of knowledge graph completion and knowledge graph error correction tasks, thereby obtaining effective extended knowledge data; The embodiment of the present invention utilizes a multi-task learning framework to screen and verify candidate extended knowledge data, so as to simultaneously optimize the completion and error correction tasks of the knowledge graph. The design of the multi-task learning framework can screen out the most reliable extended knowledge data by training different tasks (such as knowledge completion and error correction) in parallel. In specific operations, a joint training method is used to build a multi-task learning model based on a graph neural network. During the training process, the model parameters are updated through back propagation to achieve the effect of optimizing two tasks at the same time. In actual scenarios, the system may discover and correct errors in the original knowledge graph, such as incorrectly associated disease-symptom relationships, and ultimately screen out valid extended knowledge data.

[0043] Step S246: Expand the initial knowledge graph according to the effective extended knowledge data to obtain a knowledge extension graph.

[0044] The embodiment of the present invention performs graph expansion on the initial knowledge graph based on the effective extended knowledge data, and finally generates a knowledge expansion graph. In this process, the system automatically integrates the verified extended knowledge data (including newly predicted entity relationships and newly discovered knowledge) into the original knowledge graph. During operation, a graph database update tool (such as the merge function in Neo4j) is used to seamlessly add new nodes and new edges to the graph structure while maintaining the overall consistency of the graph. In practical applications, the expanded knowledge graph can more comprehensively reflect the health status and potential risks of pregnant women, helping doctors make more accurate diagnosis and treatment decisions.

[0045] The heterogeneous information network modeling of the present invention can express and organize medical knowledge more comprehensively. By constructing a heterogeneous graph structure containing multiple node types (such as patients, diseases, drugs, treatment plans, etc.) and edge types (such as causal relationships, correlation relationships, treatment effects, etc.), it can better simulate the complex relationships in the real world. This modeling method improves the richness of the knowledge graph, enabling it to more effectively support clinical decision-making and knowledge mining. The multi-view attention mechanism can dynamically adjust the weights of nodes and edges according to prior knowledge, so as to better capture the relative importance of different types of information. This process not only enhances the sensitivity of the model to key relationships, but also improves the efficiency of information integration, making the final generated weighted heterogeneous information network more recognizable and accurate, providing a stronger foundation for subsequent analysis. Representation learning through graph neural networks can effectively capture the characteristics of nodes and their structural information in the network, thereby generating low-dimensional embedding vectors. This embedding representation not only compresses the data dimension and reduces the computational complexity, but also retains rich topological information, making subsequent reasoning and analysis more efficient. Low-dimensional embedding vectors can be used for tasks such as similarity calculation and cluster analysis, improving the flexibility of data processing. By using network embedding representation to perform reasoning and prediction of entity relationships, potential knowledge extension data can be automatically identified. This process can reveal relationships that are not explicitly expressed and enrich the content of the knowledge graph. The candidate extended knowledge data obtained through reasoning provides an important basis for supplementing and improving the knowledge graph, which helps to improve the comprehensiveness and practicality of the knowledge graph. The multi-task learning framework can promote each other in the knowledge graph completion and error correction tasks by sharing knowledge and features, thereby improving the overall performance of the model. This process ensures that the added knowledge data is of high quality and validity by screening and verifying the candidate extended knowledge data, reduces the negative impact of erroneous information on the knowledge graph, and enhances the reliability of the knowledge graph. Integrating effective extended knowledge data into the initial knowledge graph can significantly improve the coverage and information content of the knowledge graph. This expansion process enables the knowledge graph to better reflect the complex relationships and diversity in actual medical scenarios, improves the intelligence level of the system, and provides more comprehensive support for clinical decision-making.

[0046] Preferably, step S3 comprises the following steps: Step S31: extracting medical entities and their associated information based on the individual information of pregnant women contained in the multimodal knowledge graph, thereby constructing an individual feature vector of the expectant pregnant woman; The embodiment of the present invention extracts relevant medical entities and their associated information based on the individual information of pregnant women in the multimodal knowledge graph, and constructs an individual feature vector of the expectant pregnant woman. In specific operations, the system traverses the nodes and edges in the knowledge graph, extracts relevant medical entities representing the expectant pregnant woman, such as medical history information, laboratory test results, ultrasound image information, fetal heart monitoring data, etc., and maps the attributes of these entities to feature values. Then, these features are vectorized to form a feature vector that characterizes the individual health status of the pregnant woman. For example, hypertension information in the medical history can be mapped to a binary feature, and heart rate variability data in fetal heart monitoring can be mapped to a continuous feature value.

[0047] Step S32: Perform low-dimensional vector space mapping based on graph embedding on the individual feature vector to obtain the individual feature low-dimensional vector, wherein the low-dimensional vector space mapping specifically involves feature extraction and low-dimensional embedding of nodes in the knowledge graph through a graph neural network, and aggregation of neighborhood information of the nodes through graph convolution; After constructing the individual feature vector, the embodiment of the present invention performs low-dimensional vector space mapping based on graph embedding. This step extracts features from the nodes in the knowledge graph through a graph neural network (such as GCN), and aggregates the neighborhood information of the nodes through a graph convolution operation. In the specific operation, a graph neural network model is used to perform aggregation operations on the feature vector and neighborhood information of each node, and compress the multi-dimensional feature information into a low-dimensional vector space. For example, for the characteristics of expectant mothers, their neighborhood may include multiple nodes related to the examination results, and the information of these nodes is aggregated into a comprehensive feature vector through a convolution operation. The final low-dimensional embedding vector can retain key feature information while reducing redundant dimensions, which facilitates subsequent risk prediction.

[0048] Step S33: performing random forest-based classification training according to the individual feature low-dimensional vector, thereby constructing an initial risk prediction model, wherein the initial risk prediction model classifies and evaluates various risk factors of pregnant women by integrating multiple decision trees; After obtaining the individual feature low-dimensional vector, the embodiment of the present invention uses the random forest algorithm for classification training to construct an initial risk prediction model. Random forest performs classification and evaluation by constructing multiple decision trees. In specific operations, the feature low-dimensional vectors of multiple pregnant women are first used as training sets, and hundreds of decision trees are constructed using the random forest algorithm, and each tree is segmented based on different features. By integrating the results of these decision trees, the model classifies and evaluates various risk factors for pregnant women, such as estimating the risk of premature birth, high-risk pregnancy or fetal distress. In practical applications, the parameters used may include 100 decision trees, with a maximum depth limit of 10 layers to balance the complexity of the model and the accuracy of prediction.

[0049] Step S34: designing an attention mechanism layer based on a graph attention network according to the multimodal knowledge graph, dynamically adjusting the feature importance of the initial risk prediction model, thereby obtaining a weighted feature vector; According to the multimodal knowledge graph, the embodiment of the present invention designs an attention mechanism layer based on the graph attention network (GAT) to dynamically adjust the feature importance of the initial risk prediction model. In the specific operation, the input individual feature vector is first processed using the graph attention network. The attention mechanism can adaptively learn the importance of each feature and assign different weights according to the contextual information in the graph. For example, for a pregnant woman, certain risk factors in the medical history may be given higher weights, and indicators in the ultrasound image may be given secondary weights. The feature weights adjusted by the attention mechanism will be output in the form of weighted feature vectors for further optimization of the risk assessment model.

[0050] Step S35: input the weighted feature vector into the initial risk prediction model for model optimization based on supervised learning, thereby obtaining an individualized risk assessment model.

[0051] The embodiment of the present invention inputs the weighted feature vector into the initial risk prediction model, optimizes the model using supervised learning, and thus obtains an individualized risk assessment model. In specific operations, real labeled data (such as actual risk results in obstetric cases) are used as supervisory signals to perform back propagation optimization on the model and update the weight parameters in the model. Through multiple rounds of iterative training, the model will gradually improve its ability to assess individual risks. In practical applications, the hyperparameter settings in the model optimization process may include learning rate, regularization coefficient, etc. The individualized risk assessment model that is ultimately outputted can provide accurate risk prediction and assessment results for the specific circumstances of each pregnant woman.

[0052] The present invention can comprehensively consider various aspects of patient data (such as medical history, laboratory results and imaging data) by extracting maternal medical entities and their associated information from multimodal knowledge graphs to form a comprehensive individual feature vector; this feature vector can not only accurately reflect the health status of each pregnant woman, but also provide a solid data foundation for subsequent risk assessment, which is helpful to realize personalized medicine. Low-dimensional vector space mapping effectively captures the potential structural information and complex relationships in individual feature vectors through graph neural networks. The graph convolution operation allows the model to aggregate the information of neighborhood nodes, thereby enhancing the understanding of the health status of pregnant women and reducing the data dimension without losing key information. This process improves the computational efficiency and accuracy of the model, which is convenient for subsequent data analysis and risk assessment. Random forest is a powerful ensemble learning method that can improve the accuracy and robustness of classification through the combination of multiple decision trees. Using individual feature low-dimensional vectors for classification training can effectively identify various risk factors for pregnant women. This method has strong resistance to noise and overfitting, helps to establish a reliable initial risk prediction model, and provides an important basis for clinical decision-making. The attention mechanism based on the graph attention network can dynamically adjust the weight according to the importance of the feature, so as to capture key information more accurately. This mechanism enables the model to focus on the main features that affect the health of pregnant women, improving the sensitivity and accuracy of the model for risk assessment. The introduction of weighted feature vectors can effectively improve the effect of subsequent model optimization. Inputting the weighted feature vectors into the initial risk prediction model for optimization can further improve the model's ability to predict individualized risks. Through supervised learning, the model can learn more detailed risk patterns and feature relationships, making the final individualized risk assessment model more accurate and practical. This personalized risk assessment can better meet clinical needs and improve the safety and health management level of pregnant women.

[0053] Preferably, step S34 includes the following steps: Step S341: Relationship representation is performed on the multimodal knowledge graph, and a graph attention network is constructed, wherein the relationship representation specifically uses nodes and edge information in the multimodal knowledge graph to represent the individual characteristics of the expectant mother and her relationship with other related nodes; The embodiment of the present invention performs relational representation on the multimodal knowledge graph and constructs a graph attention network. In the specific operation, the nodes in the knowledge graph (such as the medical history, examination results, ultrasound images, etc. of the pregnant woman) and the relationships between them (such as the association between medical history and examination results, and the association between fetal heart monitoring and ultrasound images) are represented as nodes and edges in the network structure. For the individual characteristics of the expectant pregnant woman, it is used as the central node and connected through other medical entity nodes related to it (such as the node of hypertension history and the node of blood test results) to form a subgraph with semantic information. This relationship representation can be constrained by a preset medical field knowledge ontology model, so that the generated graph attention network can accurately reflect the association structure in the actual clinical data.

[0054] Step S342: Perform weighted calculation on the node relationships in the graph attention network to obtain an attention weight matrix, wherein the weighted calculation specifically assigns different attention weights to neighbor nodes, dynamically calculates the degree of influence of each neighbor node on the central node, and the calculation of the attention weight is based on the feature similarity of the nodes and their structural position in the graph; The embodiment of the present invention performs weighted calculation on the node relationship in the graph attention network to generate an attention weight matrix. First, for each node (such as the individual feature node of a pregnant woman), the attention weights of its neighboring nodes are dynamically calculated. These neighboring nodes may be nodes related to medical history, test results, fetal development and other information. When performing weighted calculations, the feature similarity of the nodes and their structural positions in the graph are considered. The feature similarity can be calculated by the feature vector of the node, and the structural position is determined according to the connection relationship of the node in the graph. In terms of specific algorithms, the self-attention mechanism of the graph attention network (GAT) is adopted, and neighboring nodes are assigned different weights according to their feature similarity with the central node. Through weighted calculations, the system can dynamically adjust the contribution of different features to the prediction model to ensure that important features are given higher attention weights.

[0055] Step S343: Use the self-attention mechanism to perform weighted aggregation of node features in the graph attention network based on the features of neighboring nodes whose similarity is higher than a preset threshold according to the attention weight matrix, and dynamically adjust the feature importance of the initial risk prediction model to obtain a weighted feature vector.

[0056] After generating the attention weight matrix, the embodiment of the present invention uses the self-attention mechanism to perform weighted aggregation on the node features in the graph. This process aggregates the features of the neighborhood nodes whose similarity is higher than the preset threshold, and weights them according to the attention weight, and finally generates a new weighted feature vector for the central node. For example, for the individual feature nodes of pregnant women, if the feature similarity between their medical history nodes and laboratory examination nodes is higher than the preset threshold (such as 0.8), the features of these nodes will be weighted and aggregated, and used to dynamically adjust the feature importance of the initial risk prediction model. In the application, the specific parameters of the aggregation operation may include similarity threshold settings, specific calculation methods of feature weighting factors, etc. Through this operation, the accuracy of the model in individualized risk assessment can be significantly improved, and finally a weighted feature vector is obtained to optimize the initial risk prediction model.

[0057] The present invention can fully capture the complex relationship between individual characteristics of expectant mothers and related health information by characterizing the relationships between nodes and edges in a multimodal knowledge graph. This structured representation method can help the model better understand the health status of pregnant women in various situations, extract potential influencing factors and interactions, and provide a solid foundation for subsequent analysis. The construction of the graph attention network provides an effective means for dynamically analyzing and processing these relationships. By assigning different attention weights to different neighboring nodes, the model can dynamically adjust the degree of attention to each node. This mechanism can enhance the sensitivity of the model to key factors, so that important neighborhood nodes that affect the central node (the health of expectant mothers) receive more attention. When calculating the attention weight, considering the feature similarity and the node structure position helps to capture the potential correlation between nodes, thereby improving the understanding and reasoning ability of the model. This weighted calculation can significantly improve the accuracy of subsequent risk assessment. The self-attention mechanism can extract more representative feature information by weighted aggregation of neighborhood node features whose similarity is higher than a preset threshold based on the attention weight matrix. This process not only optimizes feature selection, but also dynamically adjusts the feature importance of the initial risk prediction model, so that the model can focus more on the features that have the greatest impact on risk assessment. The final weighted feature vector will improve the model's ability to predict maternal health risks and enhance the accuracy of personalized medical decisions.

[0058] Preferably, step S4 comprises the following steps: Step S41: designing a data preprocessing module for the preset newly added relevant data according to the obstetrics domain ontology model, thereby obtaining data preprocessing module data, wherein the data preprocessing module includes word segmentation and denoising processing for text, denoising and edge detection processing for images; The present invention designs a data preprocessing module for newly added relevant data (such as newly added medical records, examination results, ultrasound images, etc.) based on the obstetrics ontology model. In text data processing, natural language processing (NLP) technology is used to segment the medical record text to remove irrelevant noise words, such as stop words, repeated information, etc., for subsequent entity recognition and relationship extraction. For image data such as ultrasound images, image denoising technology, such as Gaussian filtering, is used to remove noise interference in the image, and key information such as fetal contours and organ boundaries are extracted through edge detection algorithms (such as Canny edge detection). The processed data constitutes the data preprocessing module data and provides a basis for subsequent node matching and feature extraction.

[0059] Step S42: Perform node matching based on the data preprocessing module data and the existing multimodal knowledge graph through natural language processing technology, and extract features from the image through a convolutional neural network and associate them with entity features in the graph to obtain data fusion module data; The present invention uses natural language processing technology to match nodes between the data preprocessing module data and the existing multimodal knowledge graph. For text data, a word vector model (such as Word2Vec or BERT) is used to generate a vector representation for the preprocessed medical record text, and the semantic similarity is calculated with the existing entities in the knowledge graph to complete the matching. For ultrasound image data, a convolutional neural network (CNN) model is designed to extract features from the image, such as fetal morphology or placental position, and these image features are associated and matched with entity features in the graph (such as fetal development indicators). After this process, the data fusion module data is generated, and the text and image data are successfully integrated into the knowledge graph.

[0060] Step S43: According to the data fusion module data, the multimodal knowledge graph is adjusted in real time based on new entity addition, relationship update, attribute modification and weight redistribution, thereby obtaining graph update module data; After acquiring the data from the data fusion module, the system of the present invention adjusts the multimodal knowledge graph in real time according to the addition of new entities (such as new examination results or new fetal monitoring data). The newly added data automatically expands the graph nodes through the entity addition module of the graph, such as adding the symptoms in the new medical record as new entities to the graph. For existing nodes, the system modifies or expands their attributes according to the new data to ensure that the entities and relationships in the graph accurately reflect the latest medical data. At the same time, the preset weighting mechanism is used to adjust the relationship weights between entities to ensure that the weights of key entities and important relationships in the graph are reasonably distributed. This step generates the graph update module data to ensure that the graph remains consistent with the newly added data at any time.

[0061] Step S44: Record the knowledge graph version information after each update of the graph update module data, and track the evolution path of the graph by introducing a version control strategy to obtain the version control module data; After each graph update, the system will perform version management on the updated knowledge graph. By assigning a version number to each update and recording the evolution path of the graph, the system can trace back the changes in each version of the graph. The version control strategy includes recording the newly added entities, updated attributes, and deleted node information to ensure that the historical changes of the graph are transparent and traceable. This process achieves a complete record of the evolution of the graph through the version control module data, which facilitates the subsequent analysis and verification of a specific version of the graph.

[0062] Step S45: Integrate the data preprocessing module data, the data fusion module data, the graph updating module data and the version control module data into the data interface data of the multimodal knowledge graph; After completing data preprocessing, data fusion, graph update and version control, the system integrates the data of each module into the data interface data of the multimodal knowledge graph. The data interface contains the output data of all modules and is provided to other systems or submodules in the form of an interface. The interface is designed through RESTful API or other suitable data exchange standards, and supports real-time access and dynamic updates to ensure seamless integration of data from each module.

[0063] Step S46: Dynamically update the multimodal knowledge graph according to the data interface data and the newly added relevant data to obtain a dynamically updated knowledge graph.

[0064] The present invention uses data interface data and newly added related data to dynamically update the multimodal knowledge graph in real time. Whenever new medical records, images or laboratory results enter the system, the data interface automatically calls the preprocessing, fusion, update and version control modules to ensure that the knowledge graph can be dynamically expanded and modified based on the newly added data. After this process, the system finally generates the latest dynamically updated knowledge graph, which can reflect the latest clinical data of expectant mothers in real time and assist doctors in making decisions in the field of obstetrics.

[0065] The design of the data preprocessing module of the present invention ensures the quality and availability of input data. Through the word segmentation and denoising of the text, unnecessary noise and redundant information can be eliminated, and the accuracy and consistency of the text data can be improved. Denoising and edge detection of the image help to highlight important features and reduce interference factors in the image processing process. This preprocessing provides a clear and reliable data basis for subsequent analysis and improves the efficiency of data analysis and model training. The application of natural language processing technology makes the connection between text data and knowledge graphs closer, and data from different sources can be effectively integrated through node matching, thereby enhancing the interoperability of data. The use of convolutional neural networks can extract high-dimensional features from images, so that image information is combined with entity features in the knowledge graph to form a more comprehensive data view. The data integration of the data fusion module improves the richness and diversity of information and provides deeper insights for subsequent analysis. This process ensures the timeliness and dynamic adaptability of the knowledge graph; by adding new entities and updating relationships, the latest knowledge and discoveries can be reflected in a timely manner, ensuring that the knowledge graph is always kept up to date. The modification of attributes and the redistribution of weights can be adjusted according to the importance and relevance of the data, so that the knowledge graph is more in line with actual application scenarios. This real-time update mechanism enables the knowledge graph to remain effective in a rapidly changing environment and improves the accuracy of the decision support system. The introduction of the version control strategy makes the evolution of the knowledge graph traceable. This mechanism allows the team to review historical versions when needed to audit, restore or analyze the data. This recording and tracking not only improves the transparency of the system, but also enhances the reliability of the data, helping the development team and users understand the background and reasons for the changes in the data, and thus make more scientific decisions. Integrating the data of each module into a unified data interface data can simplify the data usage process and improve the convenience of system operation. A well-integrated data interface can facilitate different applications and users to call data when needed, thereby improving the flexibility and scalability of the system. This integration not only improves the accessibility of data, but also promotes the synergy between data. Dynamically updating the knowledge graph enables the system to adapt to new information and changes in real time, improving the real-time and accuracy of decision-making. By continuously integrating new data into the knowledge graph, the system can continue to maintain the integrity and relevance of its information. This dynamic feature ensures that in a rapidly changing medical environment, doctors and researchers can obtain the latest knowledge support to make more effective clinical decisions.

[0066] Preferably, step S5 comprises the following steps: Step S51: using an individualized risk assessment model to perform risk assessment on each version in the dynamically updated knowledge graph, and generate risk assessments at multiple time points, thereby obtaining initial risk assessment time series data; The embodiment of the present invention uses the constructed individualized risk assessment model to evaluate each version in the dynamically updated knowledge graph one by one. The system evaluates various risk factors (such as fetal abnormalities, pregnancy complications, etc.) of expectant mothers by analyzing the latest data and updated node relationships in each version. During the evaluation process, the system performs multiple risk assessments on the knowledge graph versions at different times to generate time-based assessment results. These results are presented in the form of time series, that is, the initial risk assessment time series data, which can reflect the risk dynamics that change over time.

[0067] Step S52: capturing the risk change trend based on time series analysis on the initial risk assessment time series data, thereby obtaining risk assessment time series data; After obtaining the initial risk assessment time series data, the system of the embodiment of the present invention performs time series analysis on the time series data. By using time series analysis methods such as ARIMA models or long short-term memory (LSTM) neural networks, the trend of risk factors that change over time is captured. By analyzing the changes in risk at various time points, the system can identify trends in risk growth or decline, especially changes in risk at key time nodes. For example, the system may find that the risk of a particular pregnancy complication increases significantly in the third trimester. This process generates risk assessment time series data that shows the trajectory of risk changes throughout the pregnancy.

[0068] Step S53: The attention mechanism layer based on the individualized risk assessment model ranks the risk factors at different time scales according to the risk change trend data, thereby obtaining multi-scale risk impact factor weight data; The system of the embodiment of the present invention automatically analyzes the changes of risk factors at different time scales based on the risk assessment time series data through the attention mechanism layer in the individualized risk assessment model. The attention mechanism dynamically assigns the importance weights of each risk factor according to the risk assessment results at each time point. Through this weight allocation, the system sorts the risk factors at each time point and determines which factors have the greatest impact on the overall risk at a certain moment. Through this process, the system generates multi-scale risk influencing factor weight data, which provides a basis for subsequent potential risk factor mining.

[0069] Step S54: Use graph reasoning technology to semantically expand the multi-scale risk influencing factor weight data and mine potential associated risk factors to obtain key risk factor data.

[0070] Based on the multi-scale risk influencing factor weight data obtained above, the system of the embodiment of the present invention uses graph reasoning technology to semantically expand these data. In specific operations, the system mines other risk factors that are potentially associated with current risk factors from the multimodal knowledge graph through the reasoning engine. Through semantic expansion, the system can discover hidden or not directly observed associated factors. For example, the potential association between a certain laboratory test result and certain clinical manifestations may indicate additional risks for pregnant women in the future. Ultimately, the system mines key risk factor data through this process and can provide doctors with more comprehensive decision support information.

[0071] The design of the data preprocessing module of the present invention ensures the quality and availability of the newly added data. The word segmentation and denoising processing of the text can improve the accuracy of the subsequent natural language processing, making it easier to extract key information; the denoising and edge detection processing of the image can improve the effect of image feature extraction, so that the model can better identify and associate entity features in the image. This preliminary cleaning and optimization lays a solid foundation for the subsequent processing of the data and helps to improve the overall quality of the knowledge graph. Node matching and feature extraction are the key to realizing multimodal data fusion. Natural language processing technology can effectively match text data with entities in the graph, while convolutional neural networks can extract high-dimensional features from images and associate them with text features. This process not only improves the relevance between different data sources, but also ensures the comprehensiveness and accuracy of information, providing rich information support for subsequent data fusion. Real-time updating of the multimodal knowledge graph is the key to ensuring that it remains up-to-date and relevant. By adding new entities, updating relationships and modifying attributes, the content of the knowledge graph can be continuously enriched to ensure that it can reflect the latest medical knowledge and individual characteristics. Weight redistribution can help the model pay more attention to the more important parts of the current data, thereby optimizing the application effect of the knowledge graph in medical decision-making. The version control strategy provides systematic support for the management of the knowledge graph, recording the version information of each update, so that users can track the evolution of the graph and understand the changes brought about by each update. This transparency and traceability helps to improve the credibility of the system, and can also quickly locate the problem when an error occurs or a rollback is required. Integrating the data of each module to form a unified data interface not only improves the efficiency of data processing, but also facilitates subsequent dynamic updates. This integrated data interface can simplify the data flow process, make the information transmission between modules smoother, and improve the overall performance of the system. Dynamically updating the knowledge graph can ensure that it always remains relevant and effective in the rapidly changing medical field. By processing new data in real time, the model can continuously absorb new information and promptly reflect the latest developments and changes in maternal and child health management. This flexibility and adaptability are of great significance to improving the accuracy and timeliness of medical decisions and can better serve clinical needs.

[0072] Preferably, step S52 includes the following steps: Step S521: performing time series preprocessing based on missing value interpolation, outlier detection and smoothing on the initial risk assessment time series data, thereby obtaining preprocessed time series data; The embodiment of the present invention preprocesses the initial risk assessment time series data to solve the missing values, outliers and data with large fluctuations in the time series data. The missing values ​​are filled by using the mean interpolation or K nearest neighbor interpolation method to ensure the integrity of the time series. Then, statistical methods such as Z-score or IQR are used to detect and process outliers to eliminate unreasonable abnormal data. Finally, the sliding window method or exponentially weighted moving average (EWMA) is used to smooth the time series data to eliminate the impact of short-term fluctuations, thereby generating more continuous and stable preprocessed time series data.

[0073] Step S522: performing multi-scale decomposition based on wavelet transform on the pre-processed time series data to extract features of different frequency components, thereby obtaining multi-scale time series features; The embodiment of the present invention applies wavelet transform to multi-scale decomposition of pre-processed time series data. Through wavelet transform, time series data can be decomposed into multiple subsequences of different frequency components, which correspond to the change characteristics on different time scales. These decomposed features can effectively capture the trends of fast changes and slow changes in time series data, thereby obtaining multi-scale time series features. For example, the low-frequency part may reflect long-term trends, while the high-frequency part shows short-term fluctuations.

[0074] Step S523: constructing a long short-term memory network model based on the attention mechanism according to the initial risk assessment time series data to model the multi-scale time series features, thereby obtaining a risk change pattern model; The embodiment of the present invention constructs a long short-term memory (LSTM) network model based on the multi-scale time series features extracted above, and introduces an attention mechanism to enhance the ability to identify key moments. The LSTM network can capture long-term dependencies in time series data through its memory units, and the attention mechanism can dynamically assign weights according to the characteristics of each time step to ensure that the model accurately responds to key time points in the risk change pattern. When training the model, the initial risk assessment time series data is used as input, and the parameters of the model are continuously optimized to build a model that can identify risk change patterns on different time scales.

[0075] Step S524: Time-align the output of the risk change pattern model and perform nonlinear distortion elimination processing in the time series to obtain risk change series data; The embodiment of the present invention performs a time alignment operation on the output results of the risk change pattern model to ensure that the risk assessment results at different time steps can be compared and analyzed on the same timeline. At the same time, the nonlinear distortion effect in the time series is eliminated through methods such as dynamic time warping (DTW), solving the problem of the risk changes at different times being out of sync in the time dimension. This process can generate risk change sequence data with good consistency, thereby improving the accuracy of subsequent analysis.

[0076] Step S525: extract instantaneous frequency and amplitude information of the risk change sequence data based on the Hilbert-Huang transform, so as to obtain risk assessment time series data.

[0077] Based on the risk change sequence data, the embodiment of the present invention uses the Hilbert-Huang transform (HHT) to further analyze the time series data. Through HHT, the instantaneous frequency and instantaneous amplitude information of the time series data can be extracted, which respectively reflect the speed and intensity of risk changes. These instantaneous features can more accurately capture the dynamics of risk changes in the risk assessment process, thereby obtaining the final risk assessment time series data, providing a basis for subsequent decision analysis.

[0078] The present invention can ensure the integrity and accuracy of data by interpolating missing values ​​and detecting outliers in time series data. Smoothing can eliminate noise interference and make the data more stable. Time series data after these preprocessing steps can improve the robustness of subsequent analysis, make the model more sensitive to small changes in the data, and help provide more accurate risk assessment. Wavelet transform can decompose different frequency components in time series data to capture features at different time scales. Multi-scale decomposition enables the system to simultaneously process long-term and short-term risk change information, provides multi-level feature information for further model training, and improves the timeliness and comprehensiveness of risk assessment. The long short-term memory (LSTM) network can effectively capture the long-term and short-term dependencies in time series data, and the introduction of the attention mechanism enhances the flexibility of the model in processing complex multimodal data. By modeling multi-scale time series features, the risk change pattern model can accurately identify potential risk fluctuation trends and provide reliable prediction basis for further analysis. Time alignment and elimination of nonlinear distortion help eliminate time offset and deformation in the data, thereby ensuring the consistency of data at different time points. This step improves the accuracy of time series data, making risk assessment results more accurate, especially in the processing of long-term data. The Hilbert-Huang transform (HHT) can capture the hidden nonlinear characteristics in time series data by extracting instantaneous frequency and amplitude, and is particularly suitable for the analysis of complex biological signals. Through this processing, risk assessment data can more finely reflect the subtle changes in time series data, providing more timely and accurate support for clinical decision-making.

[0079] Preferably, step S6 comprises the following steps: Step S61: performing spatiotemporal alignment on key risk factor data and risk assessment time series data, and performing time scale data synchronization, thereby obtaining time series risk data; The embodiment of the present invention performs spatiotemporal alignment of the key risk factor data and the risk assessment time series data to ensure the consistency of the two in the time dimension and the space dimension. This can be achieved through the methods of timestamp matching and spatial coordinate matching. For example, if the key risk factor data records the fetal heart rate changes within a specific time period, and the risk assessment time series data records the overall risk assessment results within the same time period, the timestamps are first compared to ensure that they can be aligned at the same time point. Next, an interpolation method (such as linear interpolation or spline interpolation) is used to synchronize the time scale to solve the problem of missing data at certain time points, thereby obtaining complete time series risk data. The final output time series risk data will contain comprehensive information on key risk factors and risk assessments to facilitate subsequent analysis.

[0080] Step S62: Generate risk change trend curves at different time stages based on the time series risk data, thereby obtaining key risk trend data; The embodiment of the present invention uses time series risk data to generate risk change trend curves at different time stages. First, select a suitable time window (such as day, week or month) to group the time series risk data. For example, you can calculate the average or maximum value of the risk factors within each week on a weekly basis to obtain the risk value for each time period. Then, use methods such as linear regression or sliding average to perform trend analysis on the risk data for each time period to generate a trend curve. These trend curves are plotted using visualization tools (such as Matplotlib or Tableau) to intuitively display the key risk change trends so that the risk change pattern over time can be observed. The key risk trend data finally obtained can be used to determine the rising or falling trend of the risk.

[0081] Step S63: Perform a general mapping of the individual characteristics of the patient based on the key risk trend data, and finally obtain risk factor-individual characteristic mapping data to assist doctors in making decisions.

[0082] The embodiment of the present invention performs a general mapping of individual characteristics of patients based on key risk trend data to form risk factor-individual characteristic mapping data. First, an individual characteristic database is established to record the basic information of each patient (such as age, medical history, pregnancy status, etc.). Then, statistical analysis or machine learning models (such as support vector machines or random forests) are used to perform association analysis on key risk trend data and individual characteristics of patients. For example, by calculating correlation coefficients or applying clustering algorithms, individual characteristics related to the changing trends of risk factors are identified. Finally, a mapping relationship between risk factors and individual characteristics is constructed to form risk factor-individual characteristic mapping data, providing decision support for doctors, helping them to better understand the impact of risk factors on different patients, and thus develop personalized medical plans.

[0083] The spatiotemporal alignment of the present invention is a key step to ensure the temporal and spatial consistency between different data sets. By synchronizing the key risk factor data and the risk assessment time series data, it can more accurately reflect the impact of risk factors on the patient's health in a specific time period. This process can eliminate the errors caused by time delays or inconsistencies between data, provide more reliable time series risk data for subsequent analysis, and improve the timeliness and accuracy of the assessment results. Generating risk change trend curves can intuitively show the changes of key risk factors over time. These curves provide doctors with a dynamic perspective to help them identify the risk status of patients at different time points and their changing trends, so that they can make clinical decisions more effectively. This visualized risk trend information helps to warn potential health risks in advance and prompt timely intervention, thereby improving the safety and treatment effect of patients. Performing a pan-mapping of individual characteristics means combining key risk factors with the patient's specific health status, medical history and physiological characteristics to form a comprehensive risk assessment model. This mapping data can provide doctors with more personalized treatment plans and help to identify the main risk factors that affect the patient's health in a targeted manner. By comprehensively considering the individual characteristics and risk factors of patients, doctors can formulate more accurate and effective intervention measures, optimize the allocation of medical resources, and improve the overall health management level of patients.

[0084] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0085] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A method for constructing an obstetric midwifery decision-making model based on knowledge graph, characterized in that: The following steps are involved: Step S1: Collect multi-dimensional information from pregnant women to obtain individualized data, where the multi-dimensional information includes medical records, laboratory test results, ultrasound images, and fetal heart rate monitoring data; standardize and structure the individualized data to generate a standardized multimodal data set; Step S2: Perform entity recognition and relationship extraction on the standardized multimodal data set according to the preset obstetrics domain ontology model to obtain an initial knowledge graph; The initial knowledge graph is expanded based on prior knowledge to obtain a multimodal knowledge graph; Step S3: constructing an individual feature vector of the expectant pregnant woman according to the multimodal knowledge graph, and performing low-dimensional vector space mapping based on graph embedding to obtain an individual feature low-dimensional vector; A risk prediction model based on random forest is trained according to the low-dimensional vector of individual characteristics, and an attention mechanism layer is added to obtain an individualized risk assessment model; Step S4: Design a text- and image-based data interface according to the obstetrics domain ontology model and the multimodal knowledge graph, and dynamically update the multimodal knowledge graph using the data interface to obtain a dynamically updated knowledge graph, wherein the data interface includes a data preprocessing module, a data fusion module, a graph update module, and a version control module; Step S5: Use the individualized risk assessment model to perform multi-version risk assessment on the dynamically updated knowledge graph, thereby obtaining risk assessment time series data; identify key risk factors through the attention mechanism layer of the individualized risk assessment model, thereby obtaining key risk factor data; Step S6: Perform key risk factor trend analysis on the risk assessment time series data based on the key risk factor data, and perform pan-mapping of the patient's individual characteristics to obtain risk factor-individual characteristic mapping data, thereby assisting doctors in making decisions.

2. The method for constructing an obstetric midwifery decision-making model based on a knowledge graph according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: collecting basic information, pregnancy history, and past medical history information from pregnant women through a unified data collection form, thereby obtaining medical record text data; Step S12: Obtaining blood and urine test results through a laboratory information system to obtain laboratory test result data; Step S13: using an integrated medical image archiving and communication system to obtain ultrasound image data, including ultrasound of fetal structure, placental position, and amniotic fluid volume; Step S14: collecting fetal heart rate and fetal heart rate variability information from the expectant pregnant woman through a fetal heart rate monitoring device, thereby obtaining fetal heart rate monitoring data; Step S15: cleaning the medical record text data, laboratory test result data, ultrasound image data, and fetal heart rate monitoring data, and merging them into individualized data; Step S16: Standardize and structure the individualized data to generate a standardized multimodal data set.

3. The method for constructing an obstetric midwifery decision-making model based on a knowledge graph according to claim 2, characterized in that: Step S2 includes the following steps: Step S21: performing medical named entity recognition on the text in the standardized multimodal dataset according to a preset obstetrics domain ontology model, thereby obtaining entity recognition result data; Step S22: extracting the relationship between medical entities from the entity recognition result data using the preset obstetrics domain ontology model, and determining the semantic relationship between entities, thereby obtaining entity relationship data; Step S23: Structuring entities and relationships according to entity recognition result data and entity relationship data, thereby constructing an initial knowledge graph, in which entities serve as nodes of the graph and relationships serve as edges of the graph; Step S24: Based on the prior knowledge in the obstetrics field, the initial knowledge graph is adaptively expanded based on a heterogeneous information network, thereby obtaining a knowledge expansion graph; Step S25: mapping the association between non-text information and graph nodes of the standardized multimodal data set according to the knowledge extension graph, thereby obtaining a multi-source data fusion graph; Step S26: Perform consistency check on the multi-source data fusion graph and perform conflict resolution to obtain a multimodal knowledge graph.

4. The method for constructing an obstetric midwifery decision model based on a knowledge graph according to claim 3, characterized in that: Step S24 includes the following steps: Step S241: performing heterogeneous information network modeling on the initial knowledge graph to obtain a heterogeneous information network model, wherein the heterogeneous information network modeling specifically involves constructing a heterogeneous graph structure including multiple node types and edge types; Step S242: designing a multi-view attention mechanism based on prior knowledge in the obstetrics field, and assigning weights to nodes and edges in the heterogeneous information network model, thereby obtaining a weighted heterogeneous information network; Step S243: performing representation learning based on a graph neural network on the weighted heterogeneous information network, and generating low-dimensional embedding vectors of the nodes, thereby obtaining a network embedding representation; Step S244: inferring and predicting entity relationships in the initial knowledge graph according to the network embedding representation, thereby obtaining candidate extended knowledge data; Step S245: Using a multi-task learning framework, the candidate extended knowledge data is screened and verified based on the simultaneous optimization of knowledge graph completion and knowledge graph error correction tasks, thereby obtaining effective extended knowledge data; Step S246: Expand the initial knowledge graph based on the effective extended knowledge data to obtain a knowledge extension graph.

5. The method for constructing an obstetric midwifery decision-making model based on a knowledge graph according to claim 4, characterized in that: Step S3 includes the following steps: Step S31: extracting medical entities and their associated information based on the individual information of pregnant women contained in the multimodal knowledge graph, thereby constructing an individual feature vector of the expectant pregnant woman; Step S32: Perform low-dimensional vector space mapping based on graph embedding on the individual feature vector to obtain the individual feature low-dimensional vector, wherein the low-dimensional vector space mapping specifically involves feature extraction and low-dimensional embedding of nodes in the knowledge graph through a graph neural network, and aggregation of neighborhood information of the nodes through graph convolution; Step S33: performing random forest-based classification training according to the individual feature low-dimensional vector, thereby constructing an initial risk prediction model, wherein the initial risk prediction model classifies and evaluates various risk factors of pregnant women by integrating multiple decision trees; Step S34: designing an attention mechanism layer based on a graph attention network according to the multimodal knowledge graph, dynamically adjusting the feature importance of the initial risk prediction model, thereby obtaining a weighted feature vector; Step S35: input the weighted feature vector into the initial risk prediction model for model optimization based on supervised learning, thereby obtaining an individualized risk assessment model.

6. The method for constructing an obstetric midwifery decision-making model based on a knowledge graph according to claim 5, characterized in that: Step S34 includes the following steps: Step S341: Relationship representation is performed on the multimodal knowledge graph, and a graph attention network is constructed, wherein the relationship representation specifically uses nodes and edge information in the multimodal knowledge graph to represent the individual characteristics of the expectant mother and her relationship with other related nodes; Step S342: Perform weighted calculation on the node relationships in the graph attention network to obtain an attention weight matrix, wherein the weighted calculation specifically assigns different attention weights to neighbor nodes, dynamically calculates the degree of influence of each neighbor node on the central node, and the calculation of the attention weight is based on the feature similarity of the nodes and their structural position in the graph; Step S343: Use the self-attention mechanism to perform weighted aggregation of node features in the graph attention network based on the features of neighboring nodes whose similarity is higher than a preset threshold according to the attention weight matrix, and dynamically adjust the feature importance of the initial risk prediction model to obtain a weighted feature vector.

7. The method for constructing an obstetric midwifery decision-making model based on a knowledge graph according to claim 6, characterized in that: Step S4 includes the following steps: Step S41: designing a data preprocessing module for the preset newly added relevant data according to the obstetrics domain ontology model, thereby obtaining data preprocessing module data, wherein the data preprocessing module includes word segmentation and denoising processing for text, denoising and edge detection processing for images; Step S42: Perform node matching based on the data preprocessing module data and the existing multimodal knowledge graph through natural language processing technology, and extract features from the image through a convolutional neural network and associate them with entity features in the graph to obtain data fusion module data; Step S43: According to the data fusion module data, the multimodal knowledge graph is adjusted in real time based on new entity addition, relationship update, attribute modification and weight redistribution, thereby obtaining graph update module data; Step S44: Record the knowledge graph version information after each update of the graph update module data, and track the evolution path of the graph by introducing a version control strategy to obtain the version control module data; Step S45: Integrate the data preprocessing module data, the data fusion module data, the graph updating module data and the version control module data into the data interface data of the multimodal knowledge graph; Step S46: Dynamically update the multimodal knowledge graph according to the data interface data and the newly added relevant data to obtain a dynamically updated knowledge graph.

8. The method for constructing an obstetric midwifery decision model based on a knowledge graph according to claim 7, characterized in that: Step S5 includes the following steps: Step S51: using an individualized risk assessment model to perform risk assessment on each version in the dynamically updated knowledge graph, and generate risk assessments at multiple time points, thereby obtaining initial risk assessment time series data; Step S52: capturing the risk change trend of the initial risk assessment time series data based on time series analysis, thereby obtaining risk assessment time series data; Step S53: The attention mechanism layer based on the individualized risk assessment model ranks the risk factors at different time scales according to the risk change trend data, thereby obtaining multi-scale risk impact factor weight data; Step S54: Use graph reasoning technology to semantically expand the multi-scale risk influencing factor weight data and mine potential associated risk factors to obtain key risk factor data.

9. The method for constructing an obstetric midwifery decision-making model based on a knowledge graph according to claim 8, characterized in that: Step S52 includes the following steps: Step S521: performing time series preprocessing based on missing value interpolation, outlier detection and smoothing on the initial risk assessment time series data, thereby obtaining preprocessed time series data; Step S522: performing multi-scale decomposition based on wavelet transform on the pre-processed time series data to extract features of different frequency components, thereby obtaining multi-scale time series features; Step S523: constructing a long short-term memory network model based on the attention mechanism according to the initial risk assessment time series data to model the multi-scale time series features, thereby obtaining a risk change pattern model; Step S524: Time-align the output of the risk change pattern model and perform nonlinear distortion elimination processing in the time series to obtain risk change series data; Step S525: extract instantaneous frequency and amplitude information of the risk change sequence data based on the Hilbert-Huang transform, so as to obtain risk assessment time series data.

10. The method for constructing an obstetric midwifery decision model based on a knowledge graph according to claim 9, characterized in that: Step S6 includes the following steps: Step S61: performing spatiotemporal alignment on key risk factor data and risk assessment time series data, and performing time scale data synchronization, thereby obtaining time series risk data; Step S62: Generate risk change trend curves at different time stages based on the time series risk data, thereby obtaining key risk trend data; Step S63: Perform a general mapping of the individual characteristics of the patient based on the key risk trend data, and finally obtain risk factor-individual characteristic mapping data to assist doctors in making decisions.

Images