A multi-index fusion-based intelligent assessment method for perinatal risk of pregnant women

Abstract

The present application relates to the technical field of intelligent medical treatment and maternal and child health, and particularly relates to a pregnant woman perinatal risk intelligent evaluation method based on multi-index fusion, which acquires clinical, behavioral, environmental and real-time physiological data through multi-source acquisition and preprocessing, dynamically calculates multi-dimensional feature weights based on analytic hierarchy process and entropy weight method, uses attention mechanism and graph neural network for multi-modal fusion and risk correlation analysis, integrates data with the help of space-time alignment engine, adopts integrated learning to construct an evaluation model, carries out four-level grading early warning of low, medium, high and extremely high risk, generates and continuously optimizes personalized intervention scheme combined with case reasoning and reinforcement learning, and completes the intelligent dynamic evaluation and management closed loop of perinatal risk. The present application realizes comprehensive data integration, personalized index dynamic adjustment and intelligent risk correlation analysis of perinatal risk evaluation, effectively solves the problems of data island, rigid index and low fusion efficiency, and improves the accuracy and timeliness of the evaluation.

Description

Technical Field

[0001] This invention relates to the fields of intelligent medical and maternal and child health technology, and in particular to an intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion. Background Technology

[0002] Perinatal medicine originated in the 1950s, focusing on the health of pregnant women and fetuses. The concept of the perinatal period was first proposed by British physicians Drs. Coultas and his wife in the 1940s, marking the birth of perinatal medicine. By the 1970s, more than 20 countries worldwide had established perinatal medicine specialties. my country's perinatal medicine development started relatively late; the establishment of the Chinese Perinatal Medicine Research Association in 1979 marked the formal beginning of the discipline in my country. The development of perinatal medicine has gone through three stages: early exploration, technological innovation, and comprehensive development. From emphasizing maternal and neonatal health in the early 20th century, to entering a new stage in the mid-20th century with the development of medical imaging and fetal monitoring technologies, and then to the present day, it has become an interdisciplinary field encompassing genetics, nutrition, psychology, and other disciplines, with a significant decline in global neonatal and maternal mortality rates.

[0003] However, existing perinatal risk assessment technologies for pregnant women still have significant shortcomings. These are mainly reflected in the following aspects: In terms of data collection and processing, the data sources are limited, relying primarily on internal hospital data without integrating multi-source data. This results in low data fusion efficiency, difficulty in integrating heterogeneous multi-source data, inadequate data preprocessing techniques, and difficulty in guaranteeing data quality. Regarding risk assessment models, the assessment indicators are too simplistic to comprehensively reflect model performance under complex conditions. Real-time performance is insufficient, failing to meet millisecond-level response requirements. Personalized assessment capabilities are lacking; most existing solutions use generalized risk assessment models, lacking the ability to personalize assessments based on individual differences. In terms of system implementation and application, unreasonable module division and high system coupling, poor scalability and performance bottlenecks, and safety design flaws and insufficient data protection all seriously affect the accuracy and practicality of perinatal risk assessment for pregnant women. Summary of the Invention

[0004] This invention aims to address the problems of traditional perinatal risk assessment for pregnant women, such as the single source of data, lack of dynamic adaptability of indicator weights, insufficient fusion of multi-source features, and insufficient targeting of early warning and intervention programs. It provides an intelligent assessment method that integrates multiple indicators to achieve dynamic and accurate assessment and personalized intervention of perinatal risks for pregnant women, thereby improving the scientificity and efficiency of perinatal health management.

[0005] The present invention adopts the following technical solution:

[0006] A smart assessment method for perinatal risks of pregnant women based on multi-indicator fusion includes the following steps:

[0007] Step S1: Multi-source heterogeneous data acquisition and preprocessing: clinical data is acquired through the hospital information system interface, behavioral data is collected through the user terminal application, environmental data is acquired through IoT devices and third-party platform interfaces, and physiological parameter monitoring data is collected in real time through wearable devices.

[0008] Step S2: Multi-indicator feature extraction and dynamic weight allocation. Key features are extracted from the preprocessed data, including physiological indicators, laboratory indicators, imaging features, lifestyle features, and psychological state features. The weights of each indicator are dynamically calculated based on a weight allocation model that combines the analytic hierarchy process and the entropy weight method.

[0009] Step S3: Multimodal data fusion and risk correlation analysis. Construct a multimodal feature fusion network based on attention mechanism, use graph neural network technology to capture the complex nonlinear relationship between indicators, and construct a spatiotemporal alignment engine to align data at different time granularities and spatial locations.

[0010] Step S4: Intelligent risk assessment and graded early warning. An integrated learning method is used to build a multi-level risk assessment model. Based on the assessment results, pregnant women are divided into four levels: low risk, medium risk, high risk and very high risk. Differentiated early warning strategies are formulated for different risk levels.

[0011] Step S5: Generation and optimization of personalized risk intervention plans. A knowledge base for personalized risk intervention plans is built. Personalized intervention plans are generated by combining case reasoning and rule reasoning. The intervention plans are continuously optimized based on the feedback of intervention effects through reinforcement learning algorithms.

[0012] Furthermore, in step S1, the multi-source heterogeneous data acquisition and preprocessing are performed according to the following steps:

[0013] Step S11: Obtain clinical data of pregnant women through the hospital information system interface. Clinical data includes basic demographic information, medical history, laboratory test results, and imaging test results. Collect behavioral data of pregnant women through user terminal applications. Behavioral data includes dietary records, exercise status, sleep quality, and psychological state. Obtain environmental data through IoT devices and third-party platform interfaces. Environmental data includes meteorological data, air quality data, and living environment data. Collect real-time monitoring data of physiological parameters of pregnant women through wearable devices. Physiological parameters include blood pressure, heart rate, blood oxygen saturation, and body temperature.

[0014] Step S12: Clean the collected multi-source heterogeneous data. The cleaning process includes using box plot method to detect and remove outliers, using multiple imputation method to fill missing values, using Z-score standardization method to convert indicators of different dimensions into comparable data, and using data encryption and desensitization technology to protect the privacy of pregnant women.

[0015] Furthermore, in step S2, the multi-indicator feature extraction and dynamic weight allocation are performed according to the following steps:

[0016] Step S21: Extract physiological indicators, laboratory indicators, imaging features, lifestyle features, and psychological features from the preprocessed data. Physiological indicators include blood pressure, heart rate, and body temperature. Laboratory indicators include complete blood count, urinalysis, and coagulation function. Imaging features include ultrasound examination features. Lifestyle features include dietary regularity, exercise intensity, and sleep quality. Psychological features include mood fluctuations and stress levels.

[0017] Step S22: Principal component analysis is used to reduce the dimensionality of the extracted features and eliminate the correlation between features; recursive feature elimination method is used to screen out key features that are highly correlated with perinatal risk; correlation analysis is used to remove redundant features and retain the most discriminative feature combination.

[0018] Step S23: Construct a weight allocation model based on the combination of the analytic hierarchy process (AHP) and the entropy weight method. Determine the subjective weight of each indicator through the AHP and the objective weight of each indicator through the entropy weight method. Linearly weight the subjective and objective weights to obtain the comprehensive weight. Design a dynamic weight adjustment mechanism to adaptively adjust the weights according to individual differences among pregnant women and changes in the pregnancy cycle.

[0019] Furthermore, in step S3, multimodal data fusion and risk correlation analysis are performed according to the following steps:

[0020] Step S31: Construct a multimodal feature fusion network based on the attention mechanism, use a multi-head attention mechanism to capture the correlation between different types of indicators, and fuse different types of features through attention weights;

[0021] Step S32: Construct an indicator association network using graph neural network technology, with different indicators as nodes and the relationship between indicators as edges, and capture the complex nonlinear relationship between indicators through graph convolution operation;

[0022] Step S33: Construct a spatiotemporal alignment engine, use timestamp calibration technology to align data at different time granularities, and use spatial normalization technology to standardize data at different spatial locations;

[0023] Step S34: Based on time series analysis, study the changing trends of each indicator with the gestational period and its correlation with perinatal risks. Use a long short-term memory network model to capture the temporal change characteristics of the indicators, use Granger causality analysis to study the causal relationship between the indicators, and use association rule mining to discover the association patterns between the indicators.

[0024] Furthermore, in step S4, the intelligent risk assessment and graded early warning are performed according to the following steps:

[0025] Step S41: Employ an ensemble learning approach, combining the advantages of support vector machines, random forests, and deep learning algorithms to construct a high-precision risk assessment model; design a multi-task learning framework to simultaneously predict multiple perinatal risks, including gestational hypertension, gestational diabetes, preterm birth, and postpartum depression; and use model fusion technology to integrate the prediction results of different algorithms to improve the accuracy and stability of the assessment.

[0026] Step S42: Based on the output of the risk assessment model, classify the perinatal risks of pregnant women into four levels: low risk, medium risk, high risk, and very high risk.

[0027] Differentiated early warning strategies are designed for different risk levels, including early warning thresholds, early warning frequency, and early warning methods; a multi-level early warning mechanism is constructed, including automatic system early warning, early warning by medical staff, and early warning by family members; an early warning information push mechanism is designed to push early warning information to relevant personnel through three channels: mobile application, SMS, and telephone; and an early warning feedback mechanism is constructed to record early warning response and processing results for optimizing early warning strategies.

[0028] Furthermore, in step S5, the generation and optimization of personalized risk intervention plans are performed according to the following steps:

[0029] Step S51: Build a knowledge base for personalized risk intervention programs, integrating clinical guidelines, expert experience, and evidence-based medicine.

[0030] Step S52: Using a combination of case-based reasoning and rule-based reasoning, generate personalized intervention plans for pregnant women with different risk levels. The intervention plans include medical intervention, behavioral intervention, and psychological intervention.

[0031] Step S53: Continuously optimize the intervention plan based on feedback from the intervention effect using a reinforcement learning algorithm; design intervention effect evaluation indicators, including the degree of risk improvement, intervention compliance, and satisfaction; construct a reward function and calculate the reward value based on the intervention effect evaluation indicators; further optimize the intervention plan based on the reward value using a reinforcement learning algorithm.

[0032] Step S54: Establish a multidisciplinary collaboration mechanism to promote collaboration among doctors, nurses, nutritionists, and psychological counselors to jointly develop and implement intervention programs.

[0033] Furthermore, the weight allocation model based on the combination of the analytic hierarchy process (AHP) and the entropy weight method in step S23 is constructed according to the following steps:

[0034] Step S231: Determine the subjective weights of each indicator using the analytic hierarchy process. The specific process involves constructing a judgment matrix, calculating the eigenvector and the maximum eigenvalue, and performing a consistency check.

[0035] Step S232: Determine the objective weight of each indicator using the entropy weight method. The specific process is to calculate the entropy value of each indicator, calculate the difference coefficient, and determine the weight of each indicator based on the difference coefficient.

[0036] Step S233: Linearly weight the subjective weights and objective weights to obtain the comprehensive weight. The formula for calculating the comprehensive weight is:

[0037]

[0038] in, For the first The overall weight of each indicator For the first The subjective weight of each indicator, For the first The objective weight of each indicator, This is the subjective weighting coefficient, and its value range is... .

[0039] Further, in step S31, the multimodal feature fusion network based on the attention mechanism is constructed according to the following steps:

[0040] Step S311: Employ a multi-head attention mechanism to capture the correlation between different types of indicators, calculate the query vector, key vector, and value vector, and then calculate the attention score and attention weight;

[0041] Step S312: Calculate the weighted vector, concatenate the outputs of the multi-head attention mechanism, and complete feature fusion through a fully connected layer;

[0042] The formula for calculating the attention score is as follows:

[0043] ;

[0044] in, To query the vector matrix, The key vector matrix, It is a value vector matrix. Let be the dimension of the key vector. The function is a normalized exponential function. To query the product of the vector matrix and the transpose of the key vector matrix, To scale the product of the query vector matrix and the transpose of the key vector matrix. To normalize the scaled result, This is the product of the normalized result and the value vector matrix.

[0045] Furthermore, the reward function in step S53 is constructed according to the following steps:

[0046] Step S531: Design intervention effectiveness evaluation indicators, including the degree of risk improvement, intervention compliance, and satisfaction;

[0047] Step S532: Construct a reward function based on the designed evaluation indicators to calculate the reward value according to the intervention effect evaluation indicators. The reward function calculation formula is as follows:

[0048]

[0049] in, The total reward value, The reward value is the amount of risk improvement. To intervene in compliance reward values, This is the satisfaction reward value; The weighting coefficient for the reward value of risk improvement. The weighting coefficient for intervention compliance reward value, The weighting coefficient for the satisfaction reward value; and satisfying , , , .

[0050] Furthermore, in step S23, the dynamic weight adjustment mechanism is implemented by constructing a weight adjustment function. This function dynamically adjusts the weights of relevant indicators based on individual differences in maternal age, gestational age, and underlying medical conditions. The specific function is as follows:

[0051] ;

[0052] in, As an indicator In time Dynamic weights, As an indicator The initial weights, To adjust the coefficient, , The adjustment factor is time. and individual differences The function, , This refers to the total duration of the pregnancy cycle. This is a vector of individual difference factors. , Due to age factors, Due to gestational age factors, Basic disease factors, Other individual differences are also factors.

[0053] The advantages of this invention are:

[0054] This invention achieves comprehensive integration of clinical data, behavioral data, environmental data, and real-time physiological parameter monitoring data through multi-source heterogeneous data acquisition and preprocessing technology. It solves the problems of single data source and severe information silos in traditional assessment methods. This method integrates clinical data obtained from hospital information system interfaces, behavioral data collected by user terminal applications, environmental data obtained from IoT devices and third-party platform interfaces, and real-time physiological parameter monitoring data collected by wearable devices. Through advanced data cleaning and standardization technologies, it ensures data quality and consistency, provides a high-quality data foundation for risk assessment, and significantly improves the comprehensiveness and accuracy of risk assessment.

[0055] This invention achieves precise screening and personalized weight adjustment of perinatal risk assessment indicators through multi-indicator feature extraction and dynamic weight allocation technology. It overcomes the shortcomings of existing technologies where assessment indicators are fixed and difficult to adapt to individual differences. This method uses principal component analysis, recursive feature elimination and other techniques to screen key features, combines the analytic hierarchy process and entropy weight method to construct a comprehensive weight allocation model, and dynamically adjusts the weights according to the individual differences of pregnant women and changes in the pregnancy cycle, so that the assessment results are more in line with the actual situation of individuals and greatly improve the personalization and accuracy of risk assessment.

[0056] This invention constructs an intelligent risk assessment model based on attention mechanisms and graph neural networks through multimodal data fusion and risk correlation analysis techniques. It effectively solves the problems of low data fusion efficiency and difficulty in mining complex correlations in traditional methods. This method captures the correlation between different types of indicators through multi-head attention mechanism, constructs an indicator correlation network using graph neural network technology, and combines a spatiotemporal alignment engine to ensure data consistency. It realizes deep fusion and risk correlation analysis of multi-source heterogeneous data, significantly improves the accuracy and timeliness of risk assessment, and provides more reliable technical support for clinical decision-making. Attached Figure Description

[0057] The accompanying drawings, which form part of this application, are used to provide a further understanding of the application and to make other features, objects, and advantages of the application more apparent. The illustrative embodiments and descriptions of this application are used to explain the application and do not constitute an undue limitation of the application.

[0058] In the attached diagram:

[0059] Figure 1 This is a flowchart of an intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion, as described in Example 1. Detailed Implementation

[0060] The present invention will now be described in detail and specifically through specific embodiments to enable a better understanding of the invention. However, the following embodiments do not limit the scope of protection of the present invention.

[0061] Example 1

[0062] like Figure 1 As shown, a method for intelligent assessment of perinatal risks in pregnant women based on multi-indicator fusion includes the following steps:

[0063] Step S1: Multi-source heterogeneous data acquisition and preprocessing: clinical data is acquired through the hospital information system interface, behavioral data is collected through the user terminal application, environmental data is acquired through IoT devices and third-party platform interfaces, and physiological parameter monitoring data is collected in real time through wearable devices.

[0064] Step S2: Multi-indicator feature extraction and dynamic weight allocation. Key features are extracted from the preprocessed data, including physiological indicators, laboratory indicators, imaging features, lifestyle features, and psychological state features. Weights are then assigned based on a combination of the analytic hierarchy process (AHP) and entropy weighting.

[0065] The allocation model dynamically calculates the weights of each indicator.

[0066] Step S3: Multimodal data fusion and risk correlation analysis. Construct a multimodal feature fusion network based on attention mechanism, use graph neural network technology to capture the complex nonlinear relationship between indicators, and construct a spatiotemporal alignment engine to align data at different time granularities and spatial locations.

[0067] Step S4: Intelligent risk assessment and graded early warning. An integrated learning method is used to build a multi-level risk assessment model. Based on the assessment results, pregnant women are divided into four levels: low risk, medium risk, high risk and very high risk. Differentiated early warning strategies are formulated for different risk levels.

[0068] Step S5: Generation and optimization of personalized risk intervention plans. A knowledge base for personalized risk intervention plans is built. Personalized intervention plans are generated by combining case reasoning and rule reasoning. The intervention plans are continuously optimized based on the feedback of intervention effects through reinforcement learning algorithms.

[0069] In a specific embodiment, this method targets the entire perinatal period for pregnant women, from early pregnancy to 42 days postpartum, enabling dynamic risk assessment and intervention. Multi-source heterogeneous data collection is completed through multi-channel collaboration. The hospital information system interface connects to the hospital's HIS, LIS, and PACS systems to directly obtain clinical data verified by medical staff, ensuring data authority. The user terminal application uses a dedicated health management app for pregnant women, with a simple input interface for easy recording of daily data. IoT devices include indoor temperature and humidity sensors and air quality detectors, with third-party platform interfaces connecting to local meteorological bureaus and environmental monitoring centers to accurately acquire environmental data. Wearable devices are medical-grade wrist-worn monitoring devices that support continuous collection of physiological parameters and have a 24-hour battery life.

[0070] Multi-indicator feature extraction covers key influencing dimensions of the perinatal period. The weight model constructed by combining the analytic hierarchy process and the entropy weight method incorporates the clinical experience of obstetric experts and relies on the distribution characteristics of the data itself, reducing the bias of a single weight allocation method. The dynamic adjustment mechanism can adapt to the changes in the importance of indicators at different stages of pregnancy.

[0071] A multimodal feature fusion network based on an attention mechanism can automatically focus on features that contribute highly to risk assessment. A graph neural network uses indicators as nodes and correlation strength as edges to effectively uncover potential correlations between different types of indicators, such as physiological and environmental indicators. A spatiotemporal alignment engine addresses the temporal granularity differences between real-time monitoring data and regular prenatal checkup data, as well as spatial positioning biases from different data acquisition devices. An ensemble learning method employs a stacking strategy to fuse three algorithms. The underlying models leverage the advantages of support vector machines for handling small sample data, random forests for overfitting, and deep learning for capturing complex patterns. A four-level risk grading system sets clear definition standards based on clinical treatment guidelines, and differentiated early warning strategies adapt to the intervention needs of different risk levels. A personalized intervention plan knowledge base integrates the latest obstetric clinical guidelines and a vast number of clinical cases. Case-based reasoning and rule-based reasoning collaboratively generate initial plans, and reinforcement learning algorithms, guided by intervention effects, continuously iterate and optimize the plans to ensure adaptation to individual differences.

[0072] Furthermore, in step S1, the multi-source heterogeneous data acquisition and preprocessing are performed according to the following steps:

[0073] Step S11: Obtain clinical data of pregnant women through the hospital information system interface. Clinical data includes basic demographic information, medical history, laboratory test results, and imaging test results. Collect behavioral data of pregnant women through user terminal applications. Behavioral data includes dietary records, exercise status, sleep quality, and psychological state. Obtain environmental data through IoT devices and third-party platform interfaces. Environmental data includes meteorological data, air quality data, and living environment data. Collect real-time monitoring data of physiological parameters of pregnant women through wearable devices. Physiological parameters include blood pressure, heart rate, blood oxygen saturation, and body temperature.

[0074] Step S12: Clean the collected multi-source heterogeneous data. The cleaning process includes using box plot method to detect and remove outliers, using multiple imputation method to fill missing values, using Z-score standardization method to convert indicators of different dimensions into comparable data, and using data encryption and desensitization technology to protect the privacy of pregnant women.

[0075] In a specific embodiment, step S11, multi-source data collection, specifically covers perinatal risk-related factors. Clinical data is extracted from the hospital's HIS system using basic demographic information and medical history, from the LIS system using laboratory test results such as blood routine and urine routine tests, and from the PACS system using imaging examination results such as ultrasound and fetal heart rate monitoring. All data includes a collection timestamp and a medical staff review mark. Behavioral data is collected through a dedicated module in the health management APP. Dietary records support food category entry and food quantity estimation. Exercise data is linked to mobile phone motion sensor data for verification. Sleep quality is manually recorded through the APP and cross-checked with wearable device sleep monitoring data. Psychological state is guided by standardized scales to guide pregnant women to conduct regular assessments.

[0076] In the environmental data, meteorological data includes parameters such as daily average temperature and precipitation; air quality data focuses on PM2.5, PM10, and harmful gas concentrations; and residential environment data is collected through IoT devices, including information on indoor temperature and humidity, and ventilation duration. Physiological parameter collection utilizes medical-certified wearable devices to simultaneously collect blood pressure, heart rate, blood oxygen saturation, and body temperature data. The sampling frequency is set to once every 5 minutes, with abnormal values ​​automatically triggering higher-frequency sampling.

[0077] The data cleaning process in step S12 is executed sequentially. The box plot method detects outliers for continuous physiological parameters such as blood pressure and heart rate, and removes extreme data that exceed the normal physiological range. The multiple imputation method builds a predictive model based on datasets of pregnant women with the same gestational age and physical condition, and fills in missing data to maintain data integrity. Z-score standardization converts indicators with different units to the same numerical range, eliminating the impact of differences in indicator units on subsequent analysis. Data encryption uses the AES encryption algorithm to encrypt and store the original data. De-identification processing replaces identifying information such as names and ID numbers with anonymization, retaining only the features required for analysis such as age and gestational age, strictly adhering to relevant medical data privacy protection regulations.

[0078] Furthermore, in step S2, the multi-indicator feature extraction and dynamic weight allocation are performed according to the following steps:

[0079] Step S21: Extract physiological indicators, laboratory indicators, imaging features, lifestyle features, and psychological features from the preprocessed data. Physiological indicators include blood pressure, heart rate, and body temperature. Laboratory indicators include complete blood count, urinalysis, and coagulation function. Imaging features include ultrasound examination features. Lifestyle features include dietary regularity, exercise intensity, and sleep quality. Psychological features include mood fluctuations and stress levels.

[0080] Step S22: Principal component analysis is used to reduce the dimensionality of the extracted features and eliminate the correlation between features; recursive feature elimination method is used to screen out key features that are highly correlated with perinatal risk; correlation analysis is used to remove redundant features and retain the most discriminative feature combination.

[0081] Step S23: Construct a weight allocation model based on the combination of the analytic hierarchy process (AHP) and the entropy weight method. Determine the subjective weight of each indicator through the AHP and the objective weight of each indicator through the entropy weight method. Linearly weight the subjective and objective weights to obtain the comprehensive weight. Design a dynamic weight adjustment mechanism to adaptively adjust the weights according to individual differences among pregnant women and changes in the pregnancy cycle.

[0082] In a specific embodiment, the feature extraction in step S21 adopts an adaptation method for different data types. Physiological indicator features are extracted from time-series data collected by wearable devices, including derived features such as mean blood pressure, heart rate fluctuation amplitude, and body temperature stability, comprehensively reflecting changes in the physiological state of pregnant and postpartum women. Laboratory indicator features are extracted from core items such as white blood cell count and hemoglobin content in routine blood tests, urine protein and urine glucose levels in routine urine tests, and prothrombin time in coagulation function tests, all of which are key indicators for the clinical diagnosis of perinatal complications. Imaging features focus on parameters such as fetal biparietal diameter, placental thickness, and amniotic fluid index in ultrasound examinations, and quantitative features are automatically extracted through image recognition algorithms. Lifestyle features are obtained through statistical analysis. Dietary regularity is measured by the proportion of daily eating frequency and fixed eating times, exercise intensity is converted into metabolic equivalents (METS) by combining exercise duration and exercise type, and sleep quality is based on the proportion of deep sleep duration at night as the core indicator. Psychological state features are extracted based on scale assessment results, including features such as the frequency of mood fluctuations and stress level.

[0083] The feature optimization process in step S22 is progressive. Principal component analysis maps high-dimensional features to a low-dimensional space, retaining principal components with a cumulative contribution rate of over 85%, thus reducing the computational load of the model. Recursive feature elimination combined with cross-validation (CV) gradually removes features that have little impact on the model's prediction accuracy, selecting a subset of core features. Correlation analysis uses the Pearson correlation coefficient to remove redundant features with an absolute correlation coefficient greater than 0.7, avoiding mutual interference between features.

[0084] In step S23, during the construction of the weighted model, the analytic hierarchy process (AHP) involves inviting senior obstetrics experts to construct a judgment matrix, and then adjusting the matrix elements based on the results of the expert opinion consistency test to determine the subjective weights. The entropy weight method, based on preprocessed sample data, measures the dispersion of indicators by calculating their information entropy and determines objective weights. The overall weights are obtained through linear weighting, with α around 0.5 to balance the influence of subjective experience and objective data. The dynamic weight adjustment mechanism relies on a weight adjustment function to adapt in real time according to the individual circumstances of pregnant women. For pregnant women with a history of gestational hypertension, the weight of blood pressure-related indicators is appropriately increased; in late pregnancy, the weight of fetal-related indicators is correspondingly increased.

[0085] Furthermore, in step S3, multimodal data fusion and risk correlation analysis are performed according to the following steps:

[0086] Step S31: Construct a multimodal feature fusion network based on the attention mechanism, use a multi-head attention mechanism to capture the correlation between different types of indicators, and fuse different types of features through attention weights;

[0087] Step S32: Construct an indicator association network using graph neural network technology, with different indicators as nodes and the relationship between indicators as edges, and capture the complex nonlinear relationship between indicators through graph convolution operation;

[0088] Step S33: Construct a spatiotemporal alignment engine, use timestamp calibration technology to align data at different time granularities, and use spatial normalization technology to standardize data at different spatial locations;

[0089] Step S34: Based on time series analysis, study the changing trends of each indicator with the gestational period and its correlation with perinatal risks. Use a long short-term memory network model to capture the temporal change characteristics of the indicators, use Granger causality analysis to study the causal relationship between the indicators, and use association rule mining to discover the association patterns between the indicators.

[0090] In a specific embodiment, the multimodal feature fusion network in step S31 employs a multi-head attention mechanism with four attention heads, corresponding to five feature categories: physiological, laboratory, imaging, lifestyle, and psychological state. Each attention head focuses on capturing the correlation between one type of feature and other features. The network generates a query vector Q, a key vector K, and a value vector V through a fully connected layer. A scaling factor is introduced when calculating the attention score to avoid the gradient vanishing problem caused by excessive scaling. After the attention weights are normalized by the softmax function, the single-head attention output is obtained by weighted summation of the value vector V. The multi-head outputs are then concatenated and feature fusion is completed through a fully connected layer.

[0091] The index association network in step S32 is constructed using a graph convolutional neural network (GCN). The nodes are the extracted features, and the weights of the edges are set based on the correlation coefficients between the features. The complex nonlinear relationships between the indicators are captured through graph convolution operations.

[0092] The spatiotemporal alignment engine in step S33 uses timestamp calibration technology to align data at different time granularities to an hourly time axis; it also uses spatial normalization technology to standardize data from different spatial locations to ensure data comparability.

[0093] Step S34 uses a Long Short-Term Memory (LSTM) network to construct a time-series prediction model to capture the trend characteristics of each indicator as gestational age changes; Granger causality analysis is used to verify the causal relationship between indicators and eliminate spurious correlation interference; the association rule mining uses the Apriori algorithm to mine potential association patterns between indicators and provide a basis for risk assessment.

[0094] Furthermore, in step S4, the intelligent risk assessment and graded early warning are performed according to the following steps:

[0095] Step S41: Employ an ensemble learning approach, combining the advantages of support vector machines, random forests, and deep learning algorithms to construct a high-precision risk assessment model; design a multi-task learning framework to simultaneously predict multiple perinatal risks, including gestational hypertension, gestational diabetes, preterm birth, and postpartum depression; and use model fusion technology to integrate the prediction results of different algorithms to improve the accuracy and stability of the assessment.

[0096] Step S42: Based on the output of the risk assessment model, classify the perinatal risks of pregnant women into four levels: low risk, medium risk, high risk, and very high risk.

[0097] Differentiated early warning strategies are designed for different risk levels, including early warning thresholds, early warning frequency, and early warning methods; a multi-level early warning mechanism is constructed, including automatic system early warning, early warning by medical staff, and early warning by family members; an early warning information push mechanism is designed to push early warning information to relevant personnel through three channels: mobile application, SMS, and telephone; and an early warning feedback mechanism is constructed to record early warning response and processing results for optimizing early warning strategies.

[0098] In a specific embodiment, the ensemble learning model in step S41 adopts a two-layer architecture. The bottom layer trains three basic models: Support Vector Machine (SVM), Random Forest (RF), and Deep Learning (DL). The deep learning model uses a hybrid architecture combining Convolutional Neural Network (CNN) and LSTM to adapt to multiple types of feature inputs. The multi-task learning framework takes four perinatal risks—gestational hypertension, gestational diabetes, preterm birth, and postpartum depression—as sub-tasks. It extracts common features through a shared feature layer and sets up dedicated task layers to optimize the prediction accuracy of individual risks. The model fusion adopts a weighted voting strategy, allocating weights based on the prediction accuracy of each basic model on the validation set.

[0099] The risk level classification in step S42 is set according to clinical diagnosis and treatment standards. Low risk refers to no abnormal indicators or high-risk factors; medium risk refers to 1-2 mildly abnormal indicators and no history of high-risk diseases; high risk refers to definite abnormal indicators or 1 high-risk factor; and very high risk refers to multiple abnormal indicators or 2 or more high-risk factors. Differentiated early warning strategies are set with appropriate parameters for different levels. Low risk receives a risk alert monthly, medium risk receives a weekly alert and reminders for prenatal checkups, high risk receives daily monitoring reminders and regular follow-ups by medical staff, and very high risk receives real-time alerts and initiates emergency intervention procedures. The multi-level early warning mechanism covers automatic system alerts, medical staff alerts, and family alerts. Early warning information is prioritized for push notifications via the app; unread messages within 30 minutes trigger SMS reminders; and in emergencies, direct phone calls are made. An early warning feedback mechanism records response status and processing results for optimizing the early warning strategy.

[0100] Furthermore, in step S5, the generation and optimization of personalized risk intervention plans are performed according to the following steps:

[0101] Step S51: Build a knowledge base for personalized risk intervention programs, integrating clinical guidelines, expert experience, and evidence-based medicine.

[0102] Step S52: Using a combination of case-based reasoning and rule-based reasoning, generate personalized intervention plans for pregnant women with different risk levels. The intervention plans include medical intervention, behavioral intervention, and psychological intervention.

[0103] Step S53: Continuously optimize the intervention plan based on feedback from the intervention effect using a reinforcement learning algorithm; design intervention effect evaluation indicators, including the degree of risk improvement, intervention compliance, and satisfaction; construct a reward function and calculate the reward value based on the intervention effect evaluation indicators; further optimize the intervention plan based on the reward value using a reinforcement learning algorithm.

[0104] Step S54: Establish a multidisciplinary collaboration mechanism to promote collaboration among doctors, nurses, nutritionists, and psychological counselors to jointly develop and implement intervention programs.

[0105] In a specific embodiment, the knowledge base for personalized risk intervention in step S51 adopts a structured storage method, classifying and organizing clinical guidelines, expert experience and evidence-based medicine by risk type, gestational age, and individual physical condition, and regularly updating the latest diagnosis and treatment guidelines to ensure the timeliness of the knowledge base.

[0106] In step S52, case reasoning and rule reasoning work together. Case reasoning retrieves historical cases from the knowledge base that have the highest similarity to the characteristics of the current pregnant woman and extracts the intervention plan framework. Rule reasoning adjusts the framework based on clinical diagnosis and treatment rules to generate a personalized plan that includes medical intervention, behavioral intervention and psychological intervention.

[0107] The reinforcement learning algorithm in step S53 uses the intervention effect as a feedback signal to construct a reward function to quantify the intervention effect. The evaluation indicators include the degree of risk improvement, intervention compliance and satisfaction. The reward function guides the algorithm to iteratively optimize the intervention program parameters.

[0108] The multidisciplinary collaboration mechanism in step S54 is realized through an online consultation platform. Obstetricians coordinate the overall plan, nurses are responsible for guiding the implementation of the plan and follow-up, nutritionists provide dietary intervention suggestions, and psychological counselors provide special services for pregnant women with psychological risks. Multidisciplinary collaboration ensures the comprehensiveness of the intervention plan.

[0109] Furthermore, the weight allocation model based on the combination of the analytic hierarchy process (AHP) and the entropy weight method in step S23 is constructed according to the following steps:

[0110] Step S231: Determine the subjective weights of each indicator using the analytic hierarchy process. The specific process involves constructing a judgment matrix, calculating the eigenvector and the maximum eigenvalue, and performing a consistency check.

[0111] Step S232: Determine the objective weight of each indicator using the entropy weight method. The specific process is to calculate the entropy value of each indicator, calculate the difference coefficient, and determine the weight of each indicator based on the difference coefficient.

[0112] Step S233: Linearly weight the subjective weights and objective weights to obtain the comprehensive weight. The formula for calculating the comprehensive weight is:

[0113]

[0114] in, For the first The overall weight of each indicator For the first The subjective weight of each indicator, For the first The objective weight of each indicator, This is the subjective weighting coefficient, and its value range is... .

[0115] In a specific embodiment, the analytic hierarchy process (AHP) in step S231 constructs a three-level hierarchical structure. The target layer clarifies the core needs of perinatal risk assessment, the criteria layer divides the indicators into five categories, and the indicator layer refines each specific indicator, making the logical hierarchy of weight allocation clear. The expert team constructs a judgment matrix based on clinical experience, uses the eigenvalue method to calculate the maximum eigenvalue λmax and eigenvector, and uses the CR value to determine the consistency of expert opinions. A CR value less than 0.1 indicates the judgment matrix is ​​valid, ensuring the scientific nature of subjective weighting.

[0116] The entropy weighting method in step S232 first normalizes the index data to eliminate the influence of different dimensions. The magnitude of the entropy value reflects the dispersion of the index data. The smaller the entropy value, the stronger the distinguishing ability of the index. The difference coefficient is calculated based on the entropy value. The larger the difference coefficient, the higher the weight of the index. The allocation of objective weights depends entirely on the characteristics of the data itself.

[0117] The comprehensive weight calculation in step S233 balances the subjective weights using the α coefficient. and objective weight When there is sufficient data, the value of α should be appropriately reduced. When more objective data and expert experience are more valuable for reference, the value of α should be increased to make the allocation of comprehensive weights more in line with actual application scenarios.

[0118] Further, in step S31, the multimodal feature fusion network based on the attention mechanism is constructed according to the following steps:

[0119] Step S311: Employ a multi-head attention mechanism to capture the correlation between different types of indicators, calculate the query vector, key vector, and value vector, and then calculate the attention score and attention weight;

[0120] Step S312: Calculate the weighted vector, concatenate the outputs of the multi-head attention mechanism, and complete feature fusion through a fully connected layer;

[0121] The formula for calculating the attention score is as follows:

[0122] ;

[0123] in, To query the vector matrix, The key vector matrix, It is a value vector matrix. Let be the dimension of the key vector. The function is a normalized exponential function. To query the product of the vector matrix and the transpose of the key vector matrix, To scale the product of the query vector matrix and the transpose of the key vector matrix. To normalize the scaled result, This is the product of the normalized result and the value vector matrix.

[0124] In a specific embodiment, the multi-head attention mechanism in step S311 sets up four attention heads, each corresponding to a different feature subset, which can comprehensively capture the correlation between various features. The Q vector represents the query requirement of the current feature, the K vector is used to match the correlation information of other features, and the V vector carries the specific content of the feature. During the attention score calculation process, the transpose of Q and K is multiplied to obtain the correlation matrix, which is then divided to scale and avoid numerical overflow caused by excessively large values. The softmax function transforms the correlation matrix into a weight matrix, making the weight allocation more reasonable.

[0125] Step S312 concatenates the output of the multi-head attention mechanism by dimension and inputs it into the fully connected layer for feature fusion. The linear transformation of the fully connected layer can integrate the differentiated information captured by multi-head attention and generate a fusion feature vector with unified dimensions, providing high-quality feature input for subsequent risk association analysis.

[0126] Furthermore, the reward function in step S53 is constructed according to the following steps:

[0127] Step S531: Design intervention effectiveness evaluation indicators, including the degree of risk improvement, intervention compliance, and satisfaction;

[0128] Step S532: Construct a reward function based on the designed evaluation indicators to calculate the reward value according to the intervention effect evaluation indicators. The reward function calculation formula is as follows:

[0129]

[0130] in, The total reward value, The reward value is the amount of risk improvement. To intervene in compliance reward values, This is the satisfaction reward value; The weighting coefficient for the reward value of risk improvement. The weighting coefficient for intervention compliance reward value, The weighting coefficient for the satisfaction reward value; and satisfying , , , .

[0131] In a specific embodiment, the intervention effect evaluation indicators in step S531 are calculated quantitatively. The degree of risk improvement is determined based on the change in risk level before and after the intervention. Intervention compliance is obtained by statistical analysis of the program execution data recorded by the APP. Satisfaction is quantitatively evaluated through a postpartum follow-up questionnaire. The three indicators measure the actual effect of the intervention program from different perspectives.

[0132] In the reward function of step S532, R serves as the total reward value to guide the optimization direction of the reinforcement learning algorithm. The allocation of the three coefficients β1, β2, and β3 prioritizes the degree of risk improvement while also taking into account intervention compliance and satisfaction. This ensures that the algorithm guarantees the effectiveness of the solution while improving its feasibility and acceptance by pregnant women during the optimization process.

[0133] Furthermore, in step S23, the dynamic weight adjustment mechanism is implemented by constructing a weight adjustment function. This function dynamically adjusts the weights of relevant indicators based on individual differences in maternal age, gestational age, and underlying medical conditions. The specific function is as follows:

[0134] ;

[0135] in, As an indicator In time Dynamic weights, As an indicator The initial weights, To adjust the coefficient, , The adjustment factor is time. and individual differences The function, , This refers to the total duration of the pregnancy cycle. This is a vector of individual difference factors. , Due to age factors, Due to gestational age factors, Basic disease factors, Other individual differences are also factors.

[0136] In a specific embodiment, the core design of the dynamic weight adjustment function lies in achieving real-time adaptation of indicator weights. The dynamic weight of the indicator over time is derived from the initial weight adjustment. The adjustment coefficient is limited to a range of -1 to 1; a positive value increases the indicator weight, while a negative value decreases it, keeping the adjustment range reasonable to avoid excessive weight changes affecting the stability of the evaluation results. The adjustment factor comprehensively considers time and individual differences, covering the entire perinatal period. Different time coefficients are set for different pregnancy stages, including multiple dimensions such as age, gestational week, and underlying diseases. Each dimension has a corresponding coefficient set based on clinical experience. The time coefficient for fetal-related indicators is increased in late pregnancy, and the age coefficient for older pregnant women is correspondingly increased. This design ensures that the indicator weights always align with the individual circumstances of the pregnant woman and the characteristics of the pregnancy stage.

[0137] The specific embodiments of the present invention have been described in detail above, but they are merely examples, and the present invention is not equivalent to the specific embodiments described above. For those skilled in the art, any equivalent modifications and substitutions to the present invention are also within the scope of the present invention. Therefore, all equivalent transformations and modifications made without departing from the spirit and scope of the present invention should be covered within the scope of the present invention.

Claims

1. A method for intelligent assessment of perinatal risks in pregnant women based on multi-indicator fusion, characterized in that, Includes the following steps: Step S1: Multi-source heterogeneous data acquisition and preprocessing: clinical data is acquired through the hospital information system interface, behavioral data is collected through the user terminal application, environmental data is acquired through IoT devices and third-party platform interfaces, and physiological parameter monitoring data is collected in real time through wearable devices. Step S2: Multi-indicator feature extraction and dynamic weight allocation. Key features are extracted from the preprocessed data, including physiological indicators, laboratory indicators, imaging features, lifestyle features, and psychological state features. The weights of each indicator are dynamically calculated based on a weight allocation model that combines the analytic hierarchy process and the entropy weight method. Step S3: Multimodal data fusion and risk correlation analysis. Construct a multimodal feature fusion network based on attention mechanism, use graph neural network technology to capture the complex nonlinear relationship between indicators, and construct a spatiotemporal alignment engine to align data at different time granularities and spatial locations. Step S4: Intelligent risk assessment and graded early warning. An integrated learning method is used to build a multi-level risk assessment model. Based on the assessment results, pregnant women are divided into four levels: low risk, medium risk, high risk and very high risk. Differentiated early warning strategies are formulated for different risk levels. Step S5: Generation and optimization of personalized risk intervention plans. A knowledge base for personalized risk intervention plans is built. Personalized intervention plans are generated by combining case reasoning and rule reasoning. The intervention plans are continuously optimized based on the feedback of intervention effects through reinforcement learning algorithms.

2. The intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion according to claim 1, characterized in that, In step S1, the multi-source heterogeneous data acquisition and preprocessing are performed according to the following steps: Step S11: Obtain clinical data of pregnant women through the hospital information system interface, including basic demographic information, medical history, laboratory test results, and imaging test results; collect behavioral data of pregnant women through user terminal applications, including dietary records, exercise status, sleep quality, and psychological state; obtain environmental data through IoT devices and third-party platform interfaces, including meteorological data, air quality data, and living environment data; collect real-time monitoring data of the physiological parameters of pregnant women through wearable devices, including blood pressure, heart rate, blood oxygen saturation, and body temperature. Step S12: Clean the collected multi-source heterogeneous data. The cleaning process includes using box plot method to detect and remove outliers, using multiple imputation method to fill missing values, using Z-score standardization method to convert indicators of different dimensions into comparable data, and using data encryption and desensitization technology to protect the privacy of pregnant women.

3. The intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion according to claim 2, characterized in that, In step S2, the multi-indicator feature extraction and dynamic weight allocation are performed according to the following steps: Step S21: Extract physiological indicators, laboratory indicators, imaging features, lifestyle features, and psychological features from the preprocessed data. The physiological indicators include blood pressure, heart rate, and body temperature. The laboratory indicators include complete blood count, complete urine count, and coagulation function. The imaging features include ultrasound examination features. The lifestyle features include dietary regularity, exercise intensity, and sleep quality. The psychological features include mood fluctuation and stress level. Step S22: Principal component analysis is used to reduce the dimensionality of the extracted features and eliminate the correlation between features; recursive feature elimination method is used to screen out key features that are highly correlated with perinatal risk; correlation analysis is used to remove redundant features and retain the most discriminative feature combination. Step S23: Construct a weight allocation model based on the combination of the analytic hierarchy process (AHP) and the entropy weight method. Determine the subjective weight of each indicator through the AHP and the objective weight of each indicator through the entropy weight method. Linearly weight the subjective weight and the objective weight to obtain the comprehensive weight. The design incorporates a dynamic weight adjustment mechanism to adaptively adjust the weights based on individual differences among pregnant women and changes in the pregnancy cycle.

4. The intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion according to claim 3, characterized in that, In step S3, the multimodal data fusion and risk correlation analysis are performed according to the following steps: Step S31: Construct a multimodal feature fusion network based on an attention mechanism, using a multi-head attention mechanism to capture the correlation between different types of indicators, and fusing the different types of features through attention weights; Step S32: Construct an indicator association network using graph neural network technology, with different indicators as nodes and the relationship between indicators as edges, and capture the complex nonlinear relationship between indicators through graph convolution operation; Step S33: Construct a spatiotemporal alignment engine, use timestamp calibration technology to align data at different time granularities, and use spatial normalization technology to standardize data at different spatial locations; Step S34: Based on time series analysis, study the changing trends of each indicator with the gestational period and its correlation with perinatal risks. Use a long short-term memory network model to capture the temporal change characteristics of the indicators, use Granger causality analysis to study the causal relationship between the indicators, and use association rule mining to discover the association patterns between the indicators.

5. The intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion according to claim 4, characterized in that, In step S4, the intelligent risk assessment and graded early warning are performed according to the following steps: Step S41: Employ an ensemble learning approach, combining the advantages of support vector machines, random forests, and deep learning algorithms to construct a high-precision risk assessment model; design a multi-task learning framework to simultaneously predict multiple perinatal risks, including gestational hypertension, gestational diabetes, preterm birth, and postpartum depression; and use model fusion technology to integrate the prediction results of different algorithms to improve the accuracy and stability of the assessment. Step S42: Based on the output of the risk assessment model, classify the perinatal risks of pregnant women into four levels: low risk, medium risk, high risk, and very high risk. Differentiated early warning strategies are designed for different risk levels, including early warning thresholds, early warning frequency, and early warning methods; Establish a multi-level early warning mechanism, including automatic system early warning, early warning from medical staff, and early warning from family members; Design an early warning information push mechanism to push early warning information to relevant personnel through three channels: mobile application, SMS, and telephone. An early warning feedback mechanism is established to record the early warning response and processing results, which are used to optimize the early warning strategy.

6. The intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion according to claim 5, characterized in that, In step S5, the generation and optimization of the personalized risk intervention plan are performed according to the following steps: Step S51: Build a knowledge base for personalized risk intervention programs, integrating clinical guidelines, expert experience, and evidence-based medicine. Step S52: Using a combination of case-based reasoning and rule-based reasoning, generate personalized intervention plans for pregnant women with different risk levels. The intervention plans include medical intervention, behavioral intervention, and psychological intervention. Step S53: Continuously optimize the intervention plan based on the feedback of the intervention effect using a reinforcement learning algorithm; design intervention effect evaluation indicators, including the degree of risk improvement, intervention compliance, and satisfaction; construct a reward function and calculate the reward value based on the intervention effect evaluation indicators; further optimize the intervention plan based on the reward value using a reinforcement learning algorithm. Step S54: Establish a multidisciplinary collaboration mechanism to promote collaboration among doctors, nurses, nutritionists, and psychological counselors to jointly develop and implement intervention programs.

7. The intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion according to claim 6, characterized in that, The weight allocation model based on the combination of the analytic hierarchy process and the entropy weight method described in step S23 is constructed according to the following steps: Step S231: Determine the subjective weights of each indicator using the analytic hierarchy process. The specific process involves constructing a judgment matrix, calculating the eigenvector and the maximum eigenvalue, and performing a consistency check. Step S232: Determine the objective weight of each indicator using the entropy weight method. The specific process is to calculate the entropy value of each indicator, calculate the difference coefficient, and determine the weight of each indicator based on the difference coefficient. Step S233: Linearly weight the subjective weights and objective weights to obtain the comprehensive weight. The formula for calculating the comprehensive weight is: in, For the first The overall weight of each indicator For the first The subjective weight of each indicator, For the first The objective weight of each indicator, This is the subjective weighting coefficient, and its value range is... .

8. The intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion according to claim 7, characterized in that, In step S31, the attention-based multimodal feature fusion network is constructed according to the following steps: Step S311: Employ a multi-head attention mechanism to capture the correlation between different types of indicators, calculate the query vector, key vector, and value vector, and then calculate the attention score and attention weight; Step S312: Calculate the weighted vector, concatenate the outputs of the multi-head attention mechanism, and complete feature fusion through a fully connected layer; The formula for calculating the attention score is as follows: ;in, To query the vector matrix, The key vector matrix, It is a value vector matrix. Let be the dimension of the key vector. The function is a normalized exponential function. To query the product of the vector matrix and the transpose of the key vector matrix, To scale the product of the query vector matrix and the transpose of the key vector matrix. To normalize the scaled result, This is the product of the normalized result and the value vector matrix.

9. The intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion according to claim 8, characterized in that, The reward function described in step S53 is constructed according to the following steps: Step S531: Design intervention effect evaluation indicators, including the degree of risk improvement, intervention compliance, and satisfaction; Step S532: Construct a reward function based on the designed evaluation indicators to calculate the reward value according to the intervention effect evaluation indicators. The reward function calculation formula is as follows: in, The total reward value, The reward value is the amount of risk improvement. To intervene in compliance reward values, This is the satisfaction reward value; The weighting coefficient for the reward value of risk improvement. The weighting coefficient for intervention compliance reward value, The weighting coefficient for the satisfaction reward value; and satisfying , , , .

10. The intelligent assessment method for perinatal risks of pregnant women based on multi-indicator fusion according to claim 9, characterized in that, In step S23, the dynamic weight adjustment mechanism is implemented by constructing a weight adjustment function. This function dynamically adjusts the weights of relevant indicators based on individual differences in maternal age, gestational age, and underlying medical conditions. The specific function is as follows: ;in, As an indicator In time Dynamic weights, As an indicator The initial weights, To adjust the coefficient, , The adjustment factor is time. and individual differences The function, , This refers to the total duration of the pregnancy cycle. This is a vector of individual difference factors. , Due to age factors, Due to gestational age factors, Basic disease factors, Other individual differences are also factors.

Images