Multimodal data fusion and intelligent management system for birth defect prevention and control during pregnancy

Abstract

The application discloses a kind of pregnant period multimodal data fusion and intelligent management system for birth defect prevention and control, it is related to medical artificial intelligence technical field, in system, multimodal data acquisition module is used to collect the multi-source heterogeneous data related to pregnant period;Data fusion and preprocessing module is used to process multi-source heterogeneous data, form pregnant period multimodal feature representation;Multimodal risk index calculation module is used to calculate the comprehensive risk index related to birth defect based on pregnant period multimodal feature representation, according to comprehensive risk index, risk stratification identification is generated;Management scheme auxiliary generation module is used to generate the individualized pregnant health management auxiliary scheme corresponding to risk stratification identification according to risk stratification identification.The present application can provide individualized health guidance, structured follow-up plan and multidisciplinary collaborative prompt for different risk levels, make up the deficiency that management mode lags behind, intervention suggestion generation is low in degree of automation in prior art.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to a multimodal data fusion and intelligent management system for the prevention and control of birth defects during pregnancy. Background Technology

[0002] Birth defects are a significant public health issue affecting the quality of life of newborns and the overall health of the population. According to relevant monitoring data, neural tube defects, congenital heart disease, and chromosomal abnormalities account for the majority of different types of birth defects. High-risk pregnant women are a key population at high risk of birth defects. Currently, how to comprehensively, dynamically, and accurately assess the risk of birth defects and effectively manage it throughout the entire pregnancy is a pressing technical problem that needs to be solved.

[0003] Currently, existing technologies for birth defect risk assessment mainly include the following: risk assessment methods based on single indicators or linear models, such as Down syndrome screening using mid-pregnancy serum indicators and ultrasound assessment of fetal nuchal translucency thickness in early pregnancy; and individualized comprehensive judgment relying on the clinical experience of obstetricians. These methods provide a basis for risk screening to some extent.

[0004] However, the aforementioned existing technologies have significant drawbacks. First, these methods are typically based on single or limited data sources, failing to effectively integrate multi-source heterogeneous data generated during pregnancy, such as preconception health data, continuous prenatal examination data, high-dimensional medical imaging data, genetic testing data, and lifestyle data monitored in real-time by wearable devices. This results in incomplete risk assessment dimensions and insufficient information utilization. Second, management methods are relatively lagging, lacking a dynamic risk monitoring mechanism based on continuously updated data, making it difficult to achieve real-time risk tracking and early warning. Furthermore, the connection between risk assessment results and subsequent personalized management interventions is not close enough, and the level of automation and intelligence needs to be improved.

[0005] With the widespread adoption of electronic health records, advancements in medical imaging technology, declining costs of genetic testing, and the extensive use of wearable devices, the availability of multi-source heterogeneous data during pregnancy is becoming increasingly abundant. To improve the efficiency and accuracy of birth defect prevention, there is an urgent need to build an intelligent system capable of systematically integrating multimodal information, achieving dynamic risk assessment, and automatically generating management plans. Summary of the Invention

[0006] The technical problem to be solved by this invention is to address the shortcomings of existing technologies, and the following technical solution is provided: 1) In a first aspect, the present invention provides a multimodal data fusion and intelligent management system for pregnancy aimed at birth defect prevention and control, the specific technical solution of which is as follows: It includes a multimodal data acquisition module, a data fusion and preprocessing module, a multimodal risk indicator calculation module, and a management scheme auxiliary generation module; The multimodal data acquisition module is used to collect multi-source heterogeneous data related to pregnancy. The data fusion and preprocessing module is used to process multi-source heterogeneous data to form a multimodal feature representation of pregnancy. The multimodal risk index calculation module is used to: calculate comprehensive risk indicators related to birth defects based on the multimodal feature representation during pregnancy through machine learning models, and generate risk stratification labels based on the comprehensive risk indicators; The management plan generation module is used to generate personalized pregnancy health management assistance plans corresponding to the risk stratification labels.

[0007] The beneficial effects of the multimodal data fusion and intelligent management system for pregnancy aimed at birth defect prevention provided by this invention are as follows: The multimodal data acquisition module systematically collects multi-source heterogeneous data related to pregnancy, and processes this data through a data fusion and preprocessing module to form a unified multimodal feature representation of pregnancy. This process overcomes the shortcomings of existing technologies in utilizing multi-source information, achieving comprehensive integration and standardization of preconception health data, prenatal examination data, medical imaging data, genetic testing data, and lifestyle data generated by wearable devices, providing a more comprehensive and consistent data foundation for risk assessment. The multimodal risk index calculation module, based on the aforementioned multimodal feature representation of pregnancy, uses a machine learning model to calculate comprehensive risk indicators related to birth defects and generates risk stratification labels. This module can dynamically update the assessment results based on newly collected multi-source heterogeneous data, thus solving the problem of the lack of dynamic and continuous risk monitoring in existing technologies, and achieving real-time tracking and accurate quantification of risks for pregnant women. The management plan assistance generation module automatically generates corresponding personalized pregnancy health management assistance plans based on the risk stratification labels. This effectively connects risk assessment and clinical management, providing individualized health guidance, structured follow-up plans, and multidisciplinary collaborative prompts for different risk levels. It makes up for the shortcomings of existing technologies, such as lagging management methods and low automation of intervention suggestion generation, and improves the efficiency and systematicness of birth defect prevention and management.

[0008] Based on the above scheme, the multimodal data fusion and intelligent management system for pregnancy aimed at birth defect prevention and control of the present invention can be further improved as follows.

[0009] Furthermore, the multimodal risk indicator calculation module is also used to: dynamically update the comprehensive risk indicator based on newly collected multi-source heterogeneous data, and generate risk stratification identifiers based on the updated comprehensive risk indicator.

[0010] The beneficial effects of adopting the above-mentioned further approach are as follows: the multimodal risk index calculation module can dynamically update the comprehensive risk indicators related to birth defects based on newly collected multi-source heterogeneous data, and generate new risk stratification labels accordingly. This process enables continuous and real-time assessment of pregnant women's risks, overcoming the lag of traditional static assessments. Dynamic updates ensure that the risk stratification labels can promptly reflect the latest changes in the pregnant women's health status, providing real-time and accurate risk basis for the management plan auxiliary generation module. This allows personalized pregnancy health management auxiliary plans to be adjusted synchronously with risk evolution, improving the timeliness of risk monitoring and the targeting of management interventions.

[0011] Furthermore, the data fusion and preprocessing module is specifically used to: perform data cleaning, format standardization, and feature fusion processing on multi-source heterogeneous data to form a multimodal feature representation of pregnancy.

[0012] The beneficial effects of adopting the above-mentioned further solution are as follows: the data fusion and preprocessing module can perform data cleaning, format standardization, and feature fusion processing on multi-source heterogeneous data. This series of processes effectively solves the heterogeneity problems of multi-source data in terms of quality, format, and semantics, transforming the originally scattered and inconsistent information into a unified, high-quality multimodal feature representation of pregnancy. This standardized feature representation provides a reliable and consistent input for the downstream multimodal risk indicator calculation module, fundamentally ensuring the accuracy and comparability of the comprehensive risk indicator calculation, and laying a solid data foundation for the effective operation of the entire system.

[0013] Furthermore, the multi-source heterogeneous data includes preconception health data, prenatal examination data, medical imaging data, genetic testing data, and lifestyle data generated by wearable devices. Personalized prenatal health management support programs include individualized health guidance and lifestyle intervention recommendations, structured follow-up monitoring plans, and multidisciplinary collaborative management tips.

[0014] The beneficial effects of adopting the above-mentioned further approach are: it clearly defines five specific types of multi-source heterogeneous data, ensuring the comprehensiveness and multidimensionality of the information upon which risk assessment is based. Simultaneously, it specifies the three components of a personalized pregnancy health management support plan, making the management output specific and structured. The comprehensive coverage of data types provides rich evidence for generating accurate management plans, while the clear definition of the plan's components transforms comprehensive risk indicators and risk stratification markers into actionable health guidance, monitoring plans, and collaborative prompts, thereby achieving a seamless transition from comprehensive risk assessment to systematic and implementable management interventions.

[0015] 2) Secondly, the present invention also provides a method for multimodal data fusion and intelligent management during pregnancy for the prevention and control of birth defects, the specific technical solution of which is as follows: Collect multi-source heterogeneous data related to pregnancy; Multi-source heterogeneous data are processed to form a multimodal feature representation of pregnancy; Based on the multimodal feature representation during pregnancy, a comprehensive risk index related to birth defects is calculated through a machine learning model, and a risk stratification label is generated based on the comprehensive risk index. Based on the risk stratification markers, a personalized pregnancy health management support plan corresponding to the risk stratification markers is generated.

[0016] Based on the above scheme, the method for multimodal data fusion and intelligent management during pregnancy for birth defect prevention and control of the present invention can be further improved as follows.

[0017] Furthermore, it also includes: dynamically updating the comprehensive risk index based on newly collected multi-source heterogeneous data, and generating risk stratification identifiers based on the updated comprehensive risk index.

[0018] Furthermore, the multi-source heterogeneous data is processed to form a multimodal feature representation of pregnancy, including: data cleaning, format standardization and feature fusion processing of the multi-source heterogeneous data to form a multimodal feature representation of pregnancy.

[0019] Furthermore, the multi-source heterogeneous data includes preconception health data, prenatal examination data, medical imaging data, genetic testing data, and lifestyle data generated by wearable devices. Personalized prenatal health management support programs include individualized health guidance and lifestyle intervention recommendations, structured follow-up monitoring plans, and multidisciplinary collaborative management tips.

[0020] 3) In a third aspect, the present invention also provides an electronic device, the electronic device including a processor coupled to a memory, the memory storing at least one computer program, the at least one computer program being loaded and executed by the processor, so that the electronic device realizes any of the above-mentioned methods for multimodal data fusion and intelligent management during pregnancy for the prevention and control of birth defects.

[0021] 4) In a fourth aspect, the present invention also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements any of the above-mentioned methods for multimodal data fusion and intelligent management during pregnancy for the prevention and control of birth defects.

[0022] It should be noted that the beneficial effects of the technical solutions of the second to fourth aspects of the present invention and their corresponding possible implementations can be found in the above description of the technical effects of the first aspect and its corresponding possible implementations, and will not be repeated here. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below: Figure 1 This is a schematic diagram of the structure of a pregnancy multimodal data fusion and intelligent management system for birth defect prevention according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating a method for multimodal data fusion and intelligent management during pregnancy aimed at preventing birth defects, according to an embodiment of the present invention. Detailed Implementation

[0024] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0025] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0026] like Figure 1 As shown in the figure, an embodiment of the present invention provides a pregnancy multimodal data fusion and intelligent management system for birth defect prevention, comprising a multimodal data acquisition module, a data fusion and preprocessing module, a multimodal risk indicator calculation module, and a management scheme auxiliary generation module; The multimodal data acquisition module is used to collect multi-source heterogeneous data related to pregnancy, including preconception health data, prenatal examination data, medical imaging data, genetic testing data, and lifestyle data generated by wearable devices. The specific implementation process is as follows: 1) Develop or adapt dedicated data acquisition interfaces for different types of data sources, such as preconception health data, prenatal examination data, medical imaging data, genetic testing data, and lifestyle data generated by wearable devices. These interfaces can establish secure connections with hospital information systems, laboratory information systems, medical imaging archives and communication systems, gene sequencing platforms, and wearable device cloud service platforms. The interfaces support industry-standard data exchange formats, such as HL7 for medical text data, DICOM for medical imaging data, and JSON or XML for genetic testing data and wearable device data. Simultaneously, establish data transmission protocols, use Secure Sockets Layer (SSL) or Transport Layer Security (TLS) protocols for encryption, and implement integrity verification mechanisms, such as using hash functions to verify the integrity of data packets. A hash function can be represented as... ,in, Indicates the original data packet. This represents the calculated hash value, used for comparison before and after transmission to ensure that the data has not been tampered with.

[0027] 2) For pre-pregnancy health data and prenatal check-up data, batch capture or real-time reception is performed via the hospital information system's application programming interface (API). The capture frequency is set according to the prenatal check-up cycle, such as once per gestational week or per gestational month. For medical imaging data, DICOM format image files, including ultrasound, MRI, and other image sequences, are pulled or pushed from the medical imaging archive and communication system. Each file includes metadata such as acquisition time, device model, and patient identifier. For genetic testing data, raw sequencing data or analysis reports are obtained from the gene sequencing platform. The data format may include FASTQ, VCF, or structured tables, and is downloaded periodically via secure file transfer protocols or APIs. For lifestyle data generated by wearable devices, time-series data, such as heart rate, steps, and sleep status per minute or hour, is synchronized through cloud service interfaces provided by the device manufacturer, using authentication protocols such as OAuth to ensure access permissions.

[0028] 3) Collected data is immediately and automatically validated to check if the data format conforms to standards, if required fields are complete, and if values ​​are within reasonable ranges. For example, blood pressure values ​​in pregnancy checkup data must meet the following requirements: systolic blood pressure between 70 mmHg and 200 mmHg, and diastolic blood pressure between 40 mmHg and 130 mmHg. Validated data is stored in a central data warehouse and indexed by data type, pregnancy identifier, and timestamp. The storage adopts a hierarchical structure, with raw data stored in cold storage and frequently used data cached in hot storage to accelerate access. Simultaneously, a data collection log is recorded, with entries including data source, collection timestamp, data size, and validation status, for tracking and data quality auditing.

[0029] 4) Throughout the data collection process, data protection regulations such as the Health Insurance Portability and Accountability Act (HIPAA) are followed. All data transmission uses encrypted channels, and data is anonymized before storage, removing personally identifiable information such as names and ID numbers and replacing them with pseudo-randomly generated unique identifiers. Access control is based on a role-based access control model, ensuring that only authorized users, such as doctors or system administrators, can access specific data, and all access operations are logged in an audit database. Through these steps, the multimodal data acquisition module can efficiently and reliably collect multi-source heterogeneous data related to pregnancy, providing structured input for the data fusion and preprocessing modules.

[0030] Preconception health data refers to historical records and assessments related to a woman's health status prior to pregnancy. This data includes, but is not limited to, age, body mass index, medical history, family history of genetic diseases, lifestyle habits such as smoking and alcohol consumption, vaccination records, and preconception examination results such as complete blood counts, urinalysis, liver function tests, and kidney function tests. Preconception health data provides a baseline health status at the start of pregnancy, used to assess potential risk factors, such as the potential impact of chronic diseases or genetic predispositions on fetal development. This data is typically derived from electronic health records or health questionnaires obtained from preconception counseling clinics, stored in structured tables or documents, and covers a timeframe from several months to several years prior to pregnancy.

[0031] Prenatal checkup data refers to clinical and laboratory measurements obtained through regular prenatal checkups during pregnancy. This data includes the date of each checkup, gestational age, blood pressure, weight gain, fundal height, abdominal circumference, fetal heart rate, and laboratory test results such as blood type, blood glucose level, hemoglobin level, infection screening such as hepatitis B and syphilis, and urinalysis. Prenatal checkup data dynamically reflects the health status of the pregnant woman and fetus, and is used to monitor pregnancy progress and detect complications early, such as gestational hypertension or gestational diabetes. This data is primarily collected from the electronic medical record system in obstetric outpatient clinics and organized in a time-series format, with each data point associated with a specific gestational age and type of examination.

[0032] Medical imaging data refers to visual representations of fetal and maternal structures acquired during pregnancy using medical imaging techniques. This data includes ultrasound images, magnetic resonance imaging (MRI) images, and, when necessary, X-ray images. Ultrasound images provide fetal growth parameters such as head circumference, abdominal circumference, and femur length, as well as anatomical assessments of the development of structures such as the heart, brain, and spine. MRI images provide more detailed soft tissue contrasts for further evaluation of any abnormalities detected by ultrasound. Medical imaging data is stored in DICOM format and contains image pixel data and metadata such as acquisition device, parameters, and patient information. This data is exported from medical imaging archives and communication systems or ultrasound workstations, typically existing as a sequence of files, with each file corresponding to an image slice or time point.

[0033] Genetic testing data refers to the results of genetic material analysis of pregnant women, fetuses, or both through molecular biology techniques. This data includes chromosome karyotype analysis, gene chip data, and next-generation sequencing data such as whole-exome sequencing or whole-genome sequencing. Genetic testing data can identify chromosomal abnormalities such as Down syndrome, microdeletion / microduplication syndromes, and single-gene genetic diseases such as thalassemia. The data may be in the form of raw sequencing files such as FASTQ format, variant calling files such as VCF format, or structured reports. This data is obtained from the information systems of gene sequencing laboratories or third-party testing platforms, involves a large amount of sequence information, and usually requires specialized analysis tools to extract features such as variant sites and genotypes.

[0034] Lifestyle data generated by wearable devices refers to time-series data generated through continuous monitoring of a pregnant woman's daily activities and habits using smart wearable devices. This data includes physical activity indicators such as steps, distance traveled, calories burned, heart rate variability, sleep quality parameters such as sleep duration and deep sleep percentage, and environmental exposures such as temperature and altitude. Wearable devices synchronize data to a cloud server via Bluetooth or wireless network, and then transmit it to the system via an application programming interface (API). This data provides objective behavioral pattern information for assessing the impact of lifestyle on pregnancy outcomes; for example, lack of exercise may be associated with a risk of gestational diabetes. The data exists as a set of high-frequency sampling points and typically requires aggregation processing to generate daily or weekly statistical summaries.

[0035] The data fusion and preprocessing module is used to process multi-source heterogeneous data to form a multimodal feature representation of pregnancy. Specifically, it performs data cleaning, format standardization, and feature fusion processing on the multi-source heterogeneous data to form a multimodal feature representation of pregnancy. The specific implementation process is as follows: 1) The goal of data cleaning is to identify and correct errors, inconsistencies, and incompleteness in multi-source heterogeneous data. Data cleaning begins with data integrity verification, checking the existence of required fields in each record. For fields with missing values, such as missing blood glucose measurements in a pregnancy checkup, the system will handle them according to preset strategies. One strategy is to fill in the missing values ​​using the average of similar measurements taken by the same pregnant woman at close gestational weeks. This average can be calculated as follows: ,in, This indicates that the pregnant woman's first The second valid blood glucose measurement value, This represents the total number of valid values. The calculated average value is used to fill in missing locations. Another strategy is to mark missing genotypes at certain loci in the genetic testing data as "unknown" and not fill them. Next, the system performs outlier detection and processing. Statistical methods are used to identify values ​​that significantly deviate from the normal physiological range, such as consecutive peaks exceeding 200 beats per minute in heart rate data recorded by wearable devices. For such outliers, the system compares them with reasonable ranges in the medical knowledge base. If they are confirmed to be erroneous or unreliable data, they are removed and logged. Data cleaning also includes duplicate data identification and merging. For example, if two similar but not identical pre-pregnancy health data questionnaires from the same pregnant woman collected from different interfaces on the same day, the system will automatically merge them into a unique and most complete record based on timestamps, data source priority, and field completeness rules.

[0036] 2) The goal of format standardization is to transform the cleaned, multi-source, heterogeneous data into a unified time base, unit of measurement, and data structure. The first step in format standardization is timeline alignment. All data will be converted and correlated to a standard gestational week timeline starting from the first day of the last menstrual period. An ultrasound examination date from medical imaging data will be converted to the corresponding precise gestational week, represented as a floating-point number. For example, 12.3 weeks of gestation. The second aspect of format standardization is the unification of measurement units. The system mandates the conversion of all numerical data to the International System of Units (SI). For instance, blood pressure in prenatal checkup data is uniformly converted from millimeters of mercury to kilopascals, and sequence coverage depth in genetic testing data is converted from per million readings to standardized count values. Unit conversion uses a linear formula: ,in, Represents the original measurement value. This represents a fixed conversion factor. This represents the standardized value. The third aspect of format standardization is the normalization of the coding system. For all categorized text data, such as medical history in preconception health data or examination results in medical imaging data, the system maps them to a standard medical terminology coding system, such as the International Classification of Diseases (ICD) or Systematic Medical Terminology (SMT) coding, ensuring semantic consistency across different sources.

[0037] 3) The goal of feature fusion processing is to extract, construct, and integrate predictive features from standardized multi-source data, ultimately forming a structured multimodal feature representation of pregnancy. Feature fusion processing begins with feature extraction and construction. For numerical time-series data, such as fundal height values ​​in continuous pregnancy examination data, in addition to using the original measurements, derived features are calculated, such as the growth slope within a specific gestational week interval. Calculated through linear fitting: ,in, This indicates the change in palace height. This represents the corresponding gestational week difference. For medical imaging data, a pre-trained convolutional neural network model is used to automatically extract deep feature vectors from the images. ,in, The dimension of the feature vector, each This represents an abstract image feature. For variation information in genetic testing data, it is encoded as Boolean or numerical features, representing the presence or absence of a specific pathogenic variation or a risk score. Feature fusion processing then unifies feature representation. All extracted and constructed features, regardless of whether they originate from text, images, or sequences, are converted into fixed-length numerical vectors. Textual features are encoded into multi-dimensional vectors, image features are already numerical vectors, and time-series statistical features (such as mean and variance) also form vectors. The final step in feature fusion processing is feature vector concatenation. All feature vectors from preconception health data, prenatal examination data, medical imaging data, genetic testing data, and lifestyle data generated by wearable devices are concatenated in a predefined order to form a comprehensive, high-dimensional joint feature vector. This concatenation process can be formally represented as... ,in, The final multimodal feature representation of pregnancy is represented by a multidimensional vector. This represents a feature vector derived from preconception health data; This represents a feature vector derived from prenatal examination data; This represents a feature vector derived from medical image data; This represents a feature vector derived from gene testing data; A feature vector representing lifestyle data generated from wearable devices; This represents a vector concatenation operation. This vector... This is the output of the module, which fully encapsulates standardized health information for pregnant women across modalities and time dimensions.

[0038] The multimodal risk index calculation module is used to: calculate a comprehensive risk index related to birth defects based on multimodal feature representation during pregnancy using a machine learning model, and generate risk stratification labels based on the comprehensive risk index. The specific implementation process is as follows: 1) Maintain a library of pre-trained and validated machine learning models. When it's necessary to calculate the risk for a pregnant woman, the module loads a specified ensemble learning model from the library, such as a fusion model consisting of gradient boosting trees, support vector machines, and deep neural networks. This model function is denoted as... Simultaneously, the module retrieves the latest, unified multimodal feature representation of the pregnant woman's pregnancy from the database or real-time data stream, denoted as a vector. Its dimensions are ,Right now The system will use feature vectors Input to loaded machine learning model Before input, a final vector integrity check is performed to ensure that all necessary feature dimensions have valid values, preventing model calculation errors due to missing data.

[0039] 2) Loaded machine learning model For the input feature vector The model performs forward propagation and computation, analyzing the interactions between features through multiple layers of nonlinear transformations or complex combinations of decision rules, ultimately outputting a continuous probability value between 0 and 1. This probability value is the comprehensive risk indicator related to birth defects, denoted as [missing value]. The calculation process can be formally represented as: .in, This is a scalar value; the higher the value, the greater the likelihood that the model, based on all available multimodal information (including pre-pregnancy health data, prenatal examination data, medical imaging data, genetic testing data, and lifestyle data generated by wearable devices), judges the fetus to be likely to have birth defects. For example, the model might output... This indicates that the overall risk index is 15%. This index is a quantitative and continuous assessment result that integrates complex information from different data sources and modalities.

[0040] 3) The system has a set of tiered risk thresholds preset to integrate continuous comprehensive risk indicators. Mapped to discrete risk levels. These thresholds are determined based on statistical analysis models of large-scale historical cohort data and clinical expert consensus. Assume two thresholds are set. and And satisfy The generation of risk stratification identifiers follows a clear decision-making rule: if the comprehensive risk indicators... Less than the threshold ,Right now If the risk level is low, the resulting risk stratification label will be "low risk"; if the comprehensive risk indicators are... Greater than or equal to the threshold And less than the threshold ,Right now If the risk level is not specified, the resulting risk stratification label will be "medium risk"; if the comprehensive risk index is not specified, the risk level will be "medium risk". Greater than or equal to the threshold ,Right now If the risk stratification is not met, the resulting risk stratification identifier will be "high risk". This process is automated and requires no human intervention. The generated risk stratification identifier is linked to the pregnant woman's unique identifier, the calculation timestamp, and the comprehensive risk index used. The values ​​are stored together in the risk assessment results database.

[0041] 4) Calculate the comprehensive risk index The generated risk stratification identifiers are encapsulated to form a structured output message, which is transmitted in real time to the management plan auxiliary generation module, serving as the direct basis for generating personalized pregnancy health management auxiliary plans. Simultaneously, the module prepares for a dynamic update mechanism. When the data fusion and preprocessing module receives newly collected and processed multi-source heterogeneous data for the pregnant woman (e.g., new week's prenatal checkup data and lifestyle data generated by wearable devices), the system triggers a new round of feature representation updates. Based on the updated pregnancy multimodal feature representation... The module will then re-execute the above three steps to calculate the updated comprehensive risk index. and based on the same threshold and Generate updated risk stratification identifiers to enable dynamic risk assessment and monitoring.

[0042] The comprehensive risk index related to birth defects is a continuous probability value output by a machine learning model, used to quantitatively assess the overall probability of birth defects in a fetus conceived by a specific pregnant woman. This index is calculated based on multimodal feature representations during pregnancy; therefore, it does not rely on a single data source but rather integrates multidimensional information such as pre-pregnancy health, physiological changes during pregnancy, fetal imaging structure, genetic variation information, and maternal daily behavior patterns. The comprehensive risk index related to birth defects typically ranges from 0 to 1, with higher values ​​indicating higher risk. It provides healthcare professionals with an objective and quantitative global risk reference value that transcends traditional single-factor assessments.

[0043] Risk stratification labels are discrete risk level tags derived from comprehensive risk indicators related to birth defects, using predefined clinical decision thresholds. Common risk stratification labels include "low risk," "medium risk," and "high risk." The purpose of risk stratification labels is to transform continuous risk probabilities into intuitive, clinically understandable, and actionable classification information. Different levels of risk stratification labels are directly linked to differentiated management intensities and intervention strategies; for example, a "high risk" label triggers more intensive follow-up monitoring and higher-priority multidisciplinary consultations. Risk stratification labels are generated based on comprehensive risk indicators and are a crucial link connecting risk calculation with the generation of personalized management plans.

[0044] In another feasible approach, based on multimodal feature representations during pregnancy, a machine learning model is used to calculate comprehensive risk indicators related to birth defects, including: 1) The key entities and indicators from various data types included in the multimodal feature representation of pregnancy—specifically, preconception health data, prenatal examination data, medical imaging data, genetic testing data, and lifestyle data generated by wearable devices—are connected based on the relationship between medical knowledge graphs and clinical pathway definitions to construct a heterogeneous graph. In this graph, nodes represent different types of feature entities, and edges represent known medical associations between entities, thus transforming high-dimensional feature vectors into relation-rich graph structure data. ,in, Represents a set of nodes. This represents the set of edges, providing input for subsequent risk calculations based on graph structures. Specifically: The system receives the pre-formed multimodal feature representation of pregnancy from the data fusion and preprocessing module. The processing begins by extracting key entities and indicators from this feature representation. These entities and indicators originate from five types of data: age and chronic disease history entries in preconception health data; serological screening index values ​​and ultrasound measurements in prenatal examination data; image feature vectors extracted by convolutional neural networks in medical imaging data; specific variant sites and risk scores in gene testing data; and daily average heart rate and sleep duration statistics in lifestyle data generated by wearable devices. The system preloads a structured medical knowledge graph, which defines various relationships between clinical concepts, such as "disease-susceptibility genes," "examination indicators-diagnostic indications," and "lifestyle-physiological effects." Based on this knowledge graph, the algorithm automatically instantiates the extracted entities and indicators as graph nodes and establishes directed edges between nodes with medical associations according to the predefined relationship types in the graph. For example, an "advanced maternal age" node and a "chromosomal aneuploidy" node will establish an "increased risk" edge; a "higher ultrasound NT value" node and a "fetal heart malformation" node will establish a "suggestive association" edge. The final heterogeneous graph is formally represented as... The structure in which This represents the set of all nodes in the graph, where each node... It includes corresponding feature vectors; Represents the set of all edges in the graph, with each edge... It includes embedding vectors representing the relationship types. This multimodal heterogeneous graph transforms the high-dimensional tabular representation of pregnancy multimodal features into graph-structured data rich in explicit medical relationships, providing direct input for subsequent deep analysis based on graph neural networks.

[0045] 2) Input the multimodal heterogeneous graph constructed in the previous step into a hierarchical graph attention network. This network first aggregates node information within each modality to learn intra-modal feature representations; then, through a cross-modal attention mechanism, it fuses information between nodes from different modalities to generate a unified pregnancy risk graph embedding vector that comprehensively reflects multimodal relationships. This process enables deep modeling of complex heterogeneous relationships, which can be formally expressed as: Where H-GAT represents the hierarchical graph attention network operation. Specifically: The constructed multimodal heterogeneous graph As input, the hierarchical graph attention network comprises two cascaded attention aggregation layers. The first layer is intra-modal graph attention aggregation. The network defines a sub-graph attention network for each data type, such as a sub-network specifically for processing all pregnancy checkup data nodes. Within each modal sub-network, a multi-head graph attention mechanism is employed, where each node aggregates neighbor information by calculating attention coefficients with its neighboring nodes. For any two adjacent nodes belonging to the same modality... and Its attention coefficient The calculation formula is ,in, and It is a node and The initial feature vector, It is a learnable weight matrix. It is a learnable attention vector. This represents a vector concatenation operation. Represents a node The network first aggregates the neighbor set of nodes. After this layer of aggregation, each node obtains an updated feature representation that incorporates its internal modal relationships. The second layer is cross-modal graph attention aggregation. The network introduces a cross-modal attention layer, which allows information transfer between nodes of different data types. For example, a genetic testing data node can compute cross-modal attention weights with a medical imaging data node, thereby fusing genetic information with imaging phenotypic information. After multiple layers of cross-modal attention aggregation, the network performs global pooling on the final representations of all nodes, generating a fixed-dimensional, unified pregnancy risk graph embedding vector. The entire process can be summarized as a function. H-GAT represents the hierarchical graph attention network operation described above.

[0046] 3) Use the pregnancy risk map embedding vector obtained in the previous step. The study uses a causal inference module to decompose the overall risk influencing birth defects into contributions from modifiable risk factors and contributions from inherent risk factors. This module learns a decoupled representation. And calculate the intervention risk value respectively. and baseline risk value The final comprehensive risk index These two parts are combined and computed with learnable weights, i.e. ,in For normalization function, and Using these as weighting parameters, a comprehensive risk index related to birth defects with clinical interpretability is obtained. Specifically: The system embeds the pregnancy risk map output from the previous step into a vector. The input is fed into a causal inference module. At the core of this module is a risk decoupler, which uses an encoder-decoder structure to process the input vector. Decomposed into two sub-vectors that minimize mutual information: the contribution encoding of modifiable risk factors. Contribution coding of inherent risk factors , that is, and The features obtained from sampling are used, and one of the optimization goals of the encoder is to maximize the discriminator's error. Subsequently, two independent lightweight multilayer perceptrons are applied to... and Map them to scalar risk values: intervention risk values and baseline risk value Ultimately, the comprehensive risk indicators associated with birth defects... The result is obtained by weighted summation of these two risk values ​​followed by normalization using the Sigmoid function. The calculation formula is as follows: .in, It is the Sigmoid normalization function, which maps the result to... interval; and These are learnable scalar weight parameters that are optimized along with other network parameters during model training to balance the contributions of the two types of risk factors to the final comprehensive risk index. This design makes the final comprehensive risk index... It has preliminary interpretability and can distinguish between risks that can be changed through medical procedures and relatively fixed background risks.

[0047] 4) Establish an adversarial continuous learning framework. When newly collected multi-source heterogeneous data is input into the system, this framework, while updating the model parameters using the new data, introduces an adversarial discriminator to prevent the forgetting of previously learned knowledge. This ensures that the machine learning model can stably absorb new information and maintain the accuracy of predictions on historical data distributions during the dynamic updating of comprehensive risk indicators, thereby maintaining the reliability and consistency of assessment performance in long-term, continuous pregnancy risk management. Specifically: After the system acquires newly collected multi-source heterogeneous data through the multimodal data acquisition module, and updates the pregnancy multimodal feature representation and corresponding heterogeneous graph after processing, the machine learning model needs to be updated to calculate the new comprehensive risk index. To prevent the model from forgetting its modeling ability for historical data distribution when adapting to new data, the system establishes an adversarial continuous learning framework. This framework includes a master risk prediction model. and an adversarial discriminator In each incremental update, the system mixes batches of new data with batches randomly sampled from historical data storage. (Main Risk Prediction Model) Using all data as input, its training objective is to minimize the prediction error of the combined risk index based on both new and old data. Meanwhile, an adversarial discriminator... It is trained to accurately determine whether an input sample comes from a new data distribution or a historical data distribution. The main risk prediction model... There is also an additional training objective: to maximize the benefit of its internal feature representation to the discriminator. The level of confusion makes it impossible for the discriminator to distinguish the data source. This adversarial process is achieved through a minimax game: ,in, It is the loss function for risk prediction. It is the discriminator's adversarial loss. These are the hyperparameters that balance the two terms. Through this framework, the model... In the process of dynamically updating parameters and optimizing the ability to predict new data using newly collected multi-source heterogeneous data, the network parameters are constrained to maintain their validity in the historical data feature space. This ensures the reliability and consistency of the comprehensive risk index calculation performance in long-term, continuous pregnancy risk management, and avoids model performance degradation caused by data distribution drift over time.

[0048] The management plan generation module is used to generate personalized pregnancy health management support plans corresponding to risk stratification identifiers. These personalized pregnancy health management support plans include individualized health guidance and lifestyle intervention recommendations, structured follow-up monitoring plans, and multidisciplinary collaborative management tips. The specific implementation process is as follows: 1) The module receives the calculated risk assessment result package from a message queue or database interface. This result package contains a core data item: the risk stratification identifier, denoted as... Its value range is {low risk, medium risk, high risk}. Simultaneously, the result package includes the pregnant woman's unique identifier and a subset of key pregnancy multimodal features, such as age, gestational age, blood pressure history, specific ultrasound soft markers, and gene carrier status. This feature set is denoted as […]. The system has a pre-installed management plan knowledge base, which identifies each risk level. A corresponding basic solution template is predefined. Basic Solution Template It is a structured document framework containing three required sections: a placeholder for individualized health guidance and lifestyle intervention recommendations, a placeholder for a structured follow-up monitoring plan, and a placeholder for multidisciplinary collaborative management suggestions. The module is categorized based on the input risk stratification identifier. It accurately matches and loads the corresponding basic solution template from the knowledge base. For example, if If the risk level is "high", then load the basic solution template specifically for high-risk situations. The template pre-sets for more frequent follow-up visits and broader requirements for multidisciplinary consultations.

[0049] 2) The basic solution template General descriptions of health guidance in China, and the set of individual characteristics of pregnant women. The data is then integrated to generate specific recommendations. The system invokes the rule engine, which stores a large number of production rules in "IF-THEN" format. The rule conditions also consider risk stratification identifiers. The conclusion section, following the specific feature values, contains the corresponding specific recommendations. For example, a rule might be: "IF..." It is 'medium risk' AND characteristic 'pre-pregnancy body mass index' THEN add a suggested entry: 'Control daily calorie intake to 1800-2000 kcal, and increase at least 150 minutes of moderate-intensity aerobic exercise per week, such as brisk walking or swimming'. Another rule might be: 'IF The rule engine adds a suggestion entry for "high risk AND genetic testing data feature 'carrying the pathogenic mutation of thalassemia'" to the true THEN condition: "Recommend that the spouse undergo allothalassemia gene testing and schedule an appointment for genetic counseling for fertility risk assessment." It then iterates through all relevant rules, collecting all matching suggestion entries to create a personalized suggestion list. Finally, the system will generate this list. The items are categorized and sorted according to categories such as nutrition, exercise, psychology, and disease management, and then populated into the basic plan template. The corresponding spacer area is used to generate individualized health guidance and lifestyle intervention recommendations.

[0050] 3) Based on risk stratification identification and current gestational week (From the feature set) (Obtain from [source]), then call the follow-up plan generator. The generator has built-in functions for standard follow-up schedules under different risk levels. This function outputs a result from the current gestational week. A sequence of future check-up appointments up to the due date, along with a recommended set of check-ups for each appointment. For example, for high-risk indicators, the function... The generated plan might include: collecting prenatal data (including blood pressure and urine protein) and fetal Doppler ultrasound blood flow monitoring every two weeks; a fetal echocardiogram must be completed between 24 and 28 weeks of gestation. The plan is presented in a timeline format. The system will then check the pregnant woman's existing medical records to avoid repeating similar completed procedures in the plan. This will generate a structured follow-up monitoring plan. This is a detailed form that includes the expected date, gestational age, required tests, purpose of the tests (e.g., screening, diagnosis, monitoring), and recommendations from the implementing agency. This form is then populated into the basic protocol template. The corresponding placeholder area.

[0051] 4) Again, based on risk stratification identification and feature set For specific medical indications (such as suspected fetal heart malformation, or a history of severe preeclampsia in a pregnant woman), a pre-defined multidisciplinary team initiation rule is retrieved from the knowledge base. The rule specifies which disciplines require expert involvement. For example, a rule might be: "IF..." The system indicates a 'high-risk' risk level, and medical imaging data suggests 'fetal ventricle enlargement.' Then, a multidisciplinary collaboration mechanism is initiated, requiring the participation of: obstetrics, fetal medicine center, pediatric neurosurgery, and radiology. Based on the matching results, the system generates a multidisciplinary collaborative management prompt list. The list not only outlines the suggested disciplines but may also include the form of collaboration (such as multidisciplinary joint clinics or remote case consultations), the suggested initial collaboration timeline (such as after a specific examination in the next structured follow-up monitoring plan), and a list of required documentation (such as all medical imaging data sequences and genetic testing reports). This list of tips... Filled into the basic scheme template The last placeholder area.

[0052] 5) In the basic scheme template Once all three placeholder areas are filled with personalized content, the system automatically integrates all parts into a complete, uniformly formatted document, which is the final personalized pregnancy health management support plan. Once the plan is generated, it can be automatically pushed to the obstetrician's review interface for final confirmation, or sent directly to the pregnant woman's application, depending on system settings. (Plan document) Related to pregnancy records and risk stratification markers Binding and storage are performed. When the multimodal risk indicator calculation module dynamically updates the risk stratification identifier, this module will automatically trigger a new round of solution generation process to dynamically adjust and update the management auxiliary solution.

[0053] The individualized health guidance and lifestyle intervention recommendations are a set of behavioral improvement and health maintenance guidelines customized based on the pregnant woman's risk stratification markers and her specific physiological, genetic, and behavioral characteristics. Unlike general pregnancy information materials, these recommendations use a rule engine to combine risk levels with individual characteristics (such as obesity, blood sugar trends, genetic carrier status, and exercise habits) to generate clearly targeted action items. These recommendations cover specific adjustments to nutritional intake, the intensity and frequency of physical activity, methods for managing psychological stress, avoiding exposure to specific environmental risk factors, and self-management guidance for existing health problems, aiming to optimize pregnancy outcomes through precise lifestyle interventions.

[0054] The structured follow-up monitoring plan is a detailed schedule for future prenatal checkups and monitoring activities, based on a timeline and risk level. Starting with the current gestational week, the plan clearly specifies the timing (specific gestational week or date), required tests (such as ultrasound screening at a specific gestational week, glucose tolerance test, and coagulation function test), and the priority and frequency of each subsequent examination, according to the clinical guidelines corresponding to the risk stratification markers. This structured follow-up monitoring plan ensures that high-risk pregnant women receive more intensive and specialized monitoring, while also providing a standardized and clear prenatal checkup pathway for low-risk pregnant women, avoiding omissions of key examinations and achieving the rational allocation and efficient use of medical resources.

[0055] The multidisciplinary collaborative management prompt is an automatically generated suggestion and explanation from the system for pregnant women with complex or high-risk factors, regarding the initiation and organization of experts from different medical specialties to participate in the diagnosis and treatment process. When the risk stratification marker or specific medical indications reach a preset threshold, the prompt will clearly list the recommended disciplines for intervention (such as maternal-fetal medicine, genetic counseling, neonatology, cardiology, nutrition, etc.), the recommended form of collaboration (such as joint outpatient clinics, case discussions), and the time window and objectives for collaboration. The multidisciplinary collaborative management prompt does not directly implement consultation scheduling, but rather provides clinicians with clear, evidence-based decision support information to promote early and orderly multidisciplinary collaboration in order to develop and implement the most comprehensive pregnancy management strategy.

[0056] Optionally, in the above technical solution, the multimodal risk index calculation module is also used to: dynamically update the comprehensive risk index based on newly collected multi-source heterogeneous data, and generate risk stratification identifiers based on the updated comprehensive risk index. The specific implementation process is as follows: 1) Deploy an event listener to continuously monitor the output interface of the data fusion and preprocessing module. When newly collected multi-source heterogeneous data for a pregnant woman already under management completes fusion and preprocessing and generates an updated pregnancy multimodal feature representation fragment, an update event will be triggered. This event carries a unique identifier for the pregnant woman. and feature vector fragments representing newly added or changed data Simultaneously, the system retrieves the pregnant woman's complete historical pregnancy multimodal feature representation from storage up to the last assessment, denoted as... And the comprehensive risk index calculated last time. and risk stratification identification The system assesses the completeness of new data, such as whether it contains key indicators from a complete prenatal checkup, to decide whether to initiate an update immediately or wait for more data.

[0057] 2) Incremental data fusion and feature representation reconstruction, with the goal of integrating historical features. With new feature fragments A complete multimodal feature representation of pregnancy that reflects the latest health status was constructed. The fusion process is not a simple splicing, but rather an intelligent merging based on data type. For time-series data, such as newly added prenatal checkup data (blood pressure, fundal height) and lifestyle data generated by wearable devices (heart rate, steps), the system will... New time series points added to At the end of the corresponding time series, the system recalculates relevant statistical features (such as the average of the most recent four weeks and the trend slope). For non-time series or one-off data, such as newly generated medical imaging data (a new ultrasound examination) or new genetic testing data (a supplementary report), the system will use... Replacement of new feature vectors in The old feature vectors of the same type. The entire fusion and reconstruction process can be formally represented as a reconstruction function. Function: . It is the reconstructed, overall pregnancy multimodal feature representation vector that represents the latest state. It is a vector representing historical features; It is a new feature fragment vector; This indicates the fusion reconstruction algorithm executed according to the above rules.

[0058] 3) Load the same trained machine learning model as in the initial evaluation phase. The latest pregnancy multimodal features obtained from the previous step are represented as follows: As input to the model. Machine learning model. right A forward propagation calculation is performed, outputting a new probability value based on all the latest information. This probability value is the updated comprehensive risk indicator related to birth defects, denoted as... The calculation process is expressed as follows: This calculation incorporates the latest physiological changes, examination results, or behavioral patterns reflected in newly acquired multi-source heterogeneous data, therefore... Perhaps compared to Whether the risk level increases, decreases, or remains stable, it objectively reflects the dynamic changes in the risk level.

[0059] 4) The system uses the same preset risk thresholds as the initial assessment. and Updated comprehensive risk indicators Classify the data. According to the decision-making rules: if... Then, an updated risk stratification identifier will be generated. "Low risk"; if ,but "Medium risk"; if ,but "High risk". (Generated) Then, the system compared it with... , and timestamp Together, this is recorded as a new version in the pregnant woman's risk assessment history log. Simultaneously, the latest comprehensive risk indicators and risk stratification markers are updated in the main database. and .

[0060] 5) The results of dynamic updates (including and The data is encapsulated into a standard message. This message is immediately pushed to the management plan generation module, serving as direct input for regenerating or adjusting personalized pregnancy health management support plans. This ensures that health management recommendations, follow-up plans, and management reminders are adjusted synchronously with changes in risk. Simultaneously, update events and key results (such as changes in risk level) are recorded in the notification system, potentially triggering alerts to clinicians to ensure timely attention to significant risk changes. At this point, a complete dynamic update process ends, and the system returns to monitoring mode, awaiting the arrival of the next batch of newly collected multi-source heterogeneous data, forming a continuous risk assessment closed loop.

[0061] The newly collected multi-source heterogeneous data refers to the latest pregnancy-related data continuously acquired through the multimodal data acquisition module after the pregnant woman has entered the system management cycle. This data falls into the same category as the initially entered data, including pre-pregnancy health data (such as newly discovered family medical history), prenatal examination data (such as the latest prenatal checkup lab results), medical imaging data (such as mid-pregnancy anomaly scans), genetic testing data (such as additional prenatal diagnostic results), and lifestyle data generated by wearable devices (such as continuous sleep monitoring records from the past week). However, these data represent new points in time or newly generated medical events. The newly collected multi-source heterogeneous data is the fundamental basis for triggering dynamic updates to the risk assessment, ensuring that the system assessment is based on the latest, continuous individual health status information, rather than remaining on historical snapshots.

[0062] The technical solution of the present invention will be further described through another embodiment. The present invention integrates multi-source heterogeneous data to construct a full-cycle, multi-modal birth defect risk prediction and management system, thereby achieving accurate risk stratification and intelligent management suggestion delivery, and improving the level of birth defect prevention. The system includes a multi-modal data acquisition module, a data fusion and preprocessing module, a multi-modal risk indicator calculation module, and a management scheme auxiliary generation module, as detailed below: 1) Multimodal Data Acquisition Module and Data Fusion and Preprocessing Module: The multimodal data acquisition module is responsible for collecting multi-type data from the preconception to delivery process, i.e., multi-source heterogeneous data, including preconception health data, prenatal examination data, medical imaging data, genetic testing data, and lifestyle data generated by wearable devices. Specifically, it includes preconception health data, preconception examination data, medical imaging data, genetic testing data, and lifestyle data generated by wearable devices. Preconception health data includes clinical structured data such as age, preconception body mass index, previous pregnancy history, chronic disease history, medication history, and infection exposure history; prenatal examination data includes serological screening results, ultrasound NT value, and fetal organ ultrasound results; medical imaging data includes fetal ultrasound images; genetic testing data includes NIPT, SNP testing, and family analysis results; and lifestyle data generated by wearable devices includes diet, exercise, smoking and drinking habits, environmental exposure risks, and physiological parameters monitored by smart devices. The data fusion and preprocessing module cleans, fills in missing values, and corrects time series data from multiple sources. It also performs natural language processing to extract information from text-based doctor's medical records, pregnancy test reports, and genetic counseling records, and extracts features from medical imaging data. Ultimately, it achieves the structuring and standardization of multi-source heterogeneous data, providing a unified input for subsequent processing.

[0063] 2) Data Fusion and Preprocessing Module and Multimodal Fusion Module: The data fusion and preprocessing module further constructs a multimodal feature representation of pregnancy. This is achieved through a multimodal fusion processing flow, which includes: a feature encoder for structured pre-pregnancy health data and prenatal examination data, generating high-dimensional vectors based on deep learning or gradient boosting decision trees; an encoder for medical text data, using a large language model in the medical field to extract textual representations of pregnancy symptoms, medical advice, and examination interpretations; a feature extraction model for medical imaging data, such as a fetal ultrasound image representation model based on Vision Transformer or convolutional neural networks; and a risk vector generator for gene testing data, converting genetic information such as mutation sites and copy number variations into risk weight vectors. The multimodal fusion module uses techniques such as Transformer, cross-attention, or graph neural networks to fuse the feature vectors from different modalities, ultimately forming a unified, high-dimensional multimodal feature representation of pregnancy, i.e., a comprehensive pregnancy risk vector.

[0064] 3) Multimodal Risk Indicator Calculation Module: This module constructs a birth defect risk prediction model based on multimodal feature representations during pregnancy to calculate a comprehensive risk index and perform stratification. This module includes a basic risk prediction model, a reinforcement learning dynamic adjustment module, and a risk stratification system. The basic risk prediction model uses multimodal feature representations during pregnancy to predict common birth defects such as neural tube defects, heart defects, and chromosomal abnormalities. The reinforcement learning dynamic adjustment module updates the model parameters and risk calculation results in real time as newly collected multi-source heterogeneous data during pregnancy is updated. The risk stratification system classifies risks into low, medium, and high risk levels based on the calculated comprehensive risk index. The stratification results are dynamically optimized based on a threshold adaptive adjustment mechanism combined with historical population data. This module can dynamically update the comprehensive risk index based on newly collected multi-source heterogeneous data and generate risk stratification labels based on the updated comprehensive risk index.

[0065] 4) Management Plan Generation Assistance Module: Based on the risk stratification identifiers provided by the multimodal risk indicator calculation module, the management plan generation assistance module automatically generates personalized pregnancy health management assistance plans. The generated plans include: individualized health guidance and lifestyle intervention suggestions, structured follow-up monitoring plans, and multidisciplinary collaborative management prompts. Individualized health guidance and lifestyle intervention suggestions include folic acid supplementation, blood sugar and blood pressure control, and personalized nutrition and exercise recommendations. The structured follow-up monitoring plan includes: recommending specific examination items and times based on risk stratification and gestational age, such as NT scan, Down syndrome screening, high-precision ultrasound, NIPT, MRI, amniocentesis, etc., forming a structured monitoring plan. The multidisciplinary collaborative management prompts indicate whether and how to initiate consultations involving multiple disciplines such as genetics, obstetrics, and ultrasound, based on the risk level. In addition, the module provides risk interpretability analysis, such as using Shapley values ​​and other interpretative models to show doctors the main risk factors affecting comprehensive risk indicators, and enabling appointment reminders and early warning pushes.

[0066] 5) The system uses real-world data for model training and cross-validation. Federated learning can be introduced to enable collaborative model training across institutions, improving model performance while protecting data privacy. A model monitoring module continuously tracks model prediction performance and corrects model biases to achieve continuous optimization and improved generalization ability.

[0067] Compared with existing birth defect risk assessment methods, the technical advantages of this invention include: ① Strong multimodal fusion capability: Simultaneously utilizing multi-source heterogeneous information such as preconception health data, prenatal examination data, medical imaging data, gene testing data, and lifestyle data generated by wearable devices, making risk assessment more comprehensive and accurate. It enables dynamic and continuous risk monitoring. After each update of newly collected multi-source heterogeneous data, the multimodal risk indicator calculation module can automatically update the comprehensive risk indicator and risk stratification identifier; ② High degree of individualization in intelligent management solutions: The management solution auxiliary generation module automatically generates personalized pregnancy health management auxiliary solutions based on risk stratification identifiers, including individualized health guidance and lifestyle intervention suggestions, structured follow-up monitoring plans, and multidisciplinary collaborative management prompts, improving clinical management efficiency; ③ Strong interpretability and clinical usability: Factor contribution analysis enhances clinicians' understanding and trust in the risk assessment results; ④ Applicable to large-scale population management: The systematic process facilitates the prevention and management of high-risk pregnant women in primary healthcare institutions, contributing to improved public health outcomes.

[0068] Let's take a 32-year-old pregnant woman who registered for the first time at 10 weeks of pregnancy as an example: 1) Data Input: The multimodal data acquisition module inputs pre-pregnancy health data (age, obstetric history, G1P0, body mass index) and prenatal examination data (early pregnancy ultrasound NT value, serological screening indicators). Simultaneously, it collects lifestyle data from wearable devices (resting heart rate and activity data provided by a smart bracelet). 2) Initial Assessment: The data fusion and preprocessing module processes the above data to form a multimodal feature representation of pregnancy. The multimodal risk indicator calculation module calculates a comprehensive risk indicator based on this feature representation. Considering factors such as age (mild risk), normal NT value, normal serological indicators, and no abnormalities in wearable device data, the initial comprehensive risk indicator is classified as "medium risk." The interpretability analysis section of the management plan auxiliary generation module shows that "age greater than 30 years" is the main risk contributing factor, but other indicators significantly reduce the overall risk. 3) Dynamic Update: At 16 weeks of pregnancy, new prenatal examination data (serological screening report) shows an abnormality in a certain indicator. Newly acquired multi-source heterogeneous data triggers the process. The data fusion and preprocessing module updates the multimodal feature representation during pregnancy, the multimodal risk index calculation module recalculates and updates the comprehensive risk index, and the risk stratification label is updated to "high risk." The interpretability analysis report highlights the strong contribution of this abnormal indicator. 4) Strategy matching and triggering: Based on the "high risk" label, the management plan auxiliary generation module immediately generates and pushes a personalized pregnancy health management auxiliary plan: ① Pushes a structured follow-up monitoring plan containing suggestions for "key fetal systemic ultrasound examination" and multidisciplinary collaborative management prompts to the doctor's workstation; ② Simultaneously pushes individualized health guidance and lifestyle intervention suggestions containing personalized reminders to the pregnant woman's application: "Based on the latest screening results, it is recommended to focus on the upcoming fetal systemic ultrasound examination, and a priority appointment has been made for you." Related health education content is also attached. Subsequently, after the pregnant woman completes the examination, new medical imaging data is entered into the system, continuing to drive the next round of risk assessment and management decisions.

[0069] like Figure 2 As shown in the figure, an embodiment of the present invention provides a method for multimodal data fusion and intelligent management during pregnancy for birth defect prevention, comprising the following steps: S1. Collect multi-source heterogeneous data related to pregnancy; S2. Process multi-source heterogeneous data to form a multimodal feature representation of pregnancy; S3. Based on the multimodal feature representation during pregnancy, a comprehensive risk index related to birth defects is calculated through a machine learning model, and a risk stratification label is generated based on the comprehensive risk index; S4. Based on the risk stratification identifier, generate a personalized pregnancy health management support plan corresponding to the risk stratification identifier.

[0070] Optionally, the above technical solution also includes: dynamically updating the comprehensive risk index based on the newly collected multi-source heterogeneous data, and generating a risk stratification identifier based on the updated comprehensive risk index.

[0071] Optionally, in the above technical solution, the multi-source heterogeneous data is processed to form a multimodal feature representation of pregnancy, including: data cleaning, format standardization and feature fusion processing of the multi-source heterogeneous data to form a multimodal feature representation of pregnancy.

[0072] Optionally, in the above technical solutions, the multi-source heterogeneous data includes preconception health data, prenatal examination data, medical imaging data, gene testing data, and lifestyle data generated by wearable devices. The personalized prenatal health management support plan includes individualized health guidance and lifestyle intervention suggestions, structured follow-up monitoring plans, and multidisciplinary collaborative management tips.

[0073] It should be noted that the beneficial effects of the multimodal data fusion and intelligent management method for birth defect prevention during pregnancy provided in the above embodiments are the same as the beneficial effects of the multimodal data fusion and intelligent management system for birth defect prevention during pregnancy described above, and will not be repeated here. Furthermore, the method and system embodiments provided in the above embodiments belong to the same concept, and their specific implementation process is detailed in the system embodiments, and will not be repeated here.

[0074] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned methods for multimodal data fusion and intelligent management during pregnancy for the prevention and control of birth defects.

[0075] An embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned methods for multimodal data fusion and intelligent management during pregnancy for the prevention and control of birth defects.

[0076] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-disclosed concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0077] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multimodal data fusion and intelligent management system for pregnancy aimed at birth defect prevention, characterized in that, It includes a multimodal data acquisition module, a data fusion and preprocessing module, a multimodal risk indicator calculation module, and a management scheme auxiliary generation module; The multimodal data acquisition module is used to: acquire multi-source heterogeneous data related to pregnancy; The data fusion and preprocessing module is used to: process the multi-source heterogeneous data to form a multimodal feature representation of pregnancy; The multimodal risk index calculation module is used to: calculate a comprehensive risk index related to birth defects based on the multimodal feature representation during pregnancy using a machine learning model, and generate a risk stratification identifier based on the comprehensive risk index; The management plan auxiliary generation module is used to generate a personalized pregnancy health management auxiliary plan corresponding to the risk stratification identifier based on the risk stratification identifier.

2. The pregnancy multimodal data fusion and intelligent management system for birth defect prevention as described in claim 1, characterized in that, The multimodal risk index calculation module is also used to: dynamically update the comprehensive risk index based on newly collected multi-source heterogeneous data, and generate risk stratification identifiers based on the updated comprehensive risk index.

3. The pregnancy multimodal data fusion and intelligent management system for birth defect prevention as described in claim 1, characterized in that, The data fusion and preprocessing module is specifically used to: perform data cleaning, format standardization and feature fusion processing on the multi-source heterogeneous data to form a multimodal feature representation of pregnancy.

4. A pregnancy multimodal data fusion and intelligent management system for birth defect prevention according to any one of claims 1 to 3, characterized in that, The multi-source heterogeneous data includes preconception health data, prenatal examination data, medical imaging data, gene testing data, and lifestyle data generated by wearable devices. The personalized prenatal health management support plan includes individualized health guidance and lifestyle intervention suggestions, structured follow-up monitoring plans, and multidisciplinary collaborative management tips.

5. A method for multimodal data fusion and intelligent management during pregnancy for birth defect prevention, characterized in that, include: Collect multi-source heterogeneous data related to pregnancy; The multi-source heterogeneous data is processed to form a multimodal feature representation of pregnancy; Based on the aforementioned multimodal features during pregnancy, a comprehensive risk index related to birth defects is calculated using a machine learning model, and a risk stratification identifier is generated based on the comprehensive risk index. Based on the risk stratification identifier, a personalized pregnancy health management support plan corresponding to the risk stratification identifier is generated.

6. The method for multimodal data fusion and intelligent management during pregnancy for birth defect prevention as described in claim 5, characterized in that, Also includes: Based on the newly collected multi-source heterogeneous data, the comprehensive risk index is dynamically updated, and a risk stratification identifier is generated based on the updated comprehensive risk index.

7. A method for multimodal data fusion and intelligent management during pregnancy for birth defect prevention as described in claim 6, characterized in that, The process of processing the multi-source heterogeneous data to form a multimodal feature representation of pregnancy includes: data cleaning, format standardization, and feature fusion processing of the multi-source heterogeneous data to form a multimodal feature representation of pregnancy.

8. A method for multimodal data fusion and intelligent management during pregnancy for birth defect prevention according to any one of claims 5 to 7, characterized in that, The multi-source heterogeneous data includes preconception health data, prenatal examination data, medical imaging data, gene testing data, and lifestyle data generated by wearable devices. The personalized prenatal health management support plan includes individualized health guidance and lifestyle intervention suggestions, structured follow-up monitoring plans, and multidisciplinary collaborative management tips.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the method for fusion and intelligent management of multimodal data during pregnancy for birth defect prevention as described in any one of claims 5 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for multimodal data fusion and intelligent management during pregnancy for birth defect prevention as described in any one of claims 5 to 8.

Images