A method for constructing a temporal multimodal preeclampsia prediction model, a prediction method, and a prediction device.

Abstract

This invention discloses a method, prediction method, and prediction device for constructing a preeclampsia prediction model based on temporal multimodal data, relating to the field of preeclampsia risk prediction technology. The construction method includes: acquiring multimodal temporal feature data of pregnant women with known pregnancy outcomes to construct a time-series dataset; the pregnant women with known pregnancy outcomes include those with and without preeclampsia; training a graph attention network and a fully connected network using the time-series dataset; embedding the resulting missing value fitting model as an embedding layer into a gated recurrent unit; and obtaining the preeclampsia prediction model after training with the time-series dataset. In this invention, the prediction model processes missing data through a missing value fitting model, effectively recovering missing information and ensuring the model can more accurately handle incomplete medical data. Simultaneously, the combination of a gated recurrent unit for capturing and predicting temporal features effectively improves the prediction accuracy of preeclampsia.

Description

Technical Field

[0001] This invention relates to the field of preeclampsia risk prediction technology, and in particular to a method for constructing a preeclampsia prediction model based on time-series multimodal analysis, a prediction method, and a prediction device. Background Technology

[0002] Preeclampsia (PE) is a multisystemic disease specific to pregnancy that poses a significant safety threat to pregnant women and perinatal infants. Currently, the diagnosis of PE mainly relies on indicators such as hypertension and proteinuria, but these methods have significant time lags and cannot meet the needs of early prediction. Furthermore, other characteristic data highly correlated with PE are often overlooked during clinical testing.

[0003] The pathogenesis of preeclampsia is not fully understood, and due to the fact that electronic health record (EHR) data is often missing and irregular, traditional time series analysis methods are difficult to process these data effectively, resulting in the inability to fully explore the potential factors and characteristics that affect the occurrence of preeclampsia and the inability to accurately predict preeclampsia.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the purpose of this invention is to provide a method for constructing a preeclampsia prediction model based on time series multimodal analysis, a prediction method and a prediction device, which aims to solve the problem that traditional time series analysis methods are difficult to effectively handle missing data, resulting in inaccurate preeclampsia prediction.

[0006] The technical solution of the present invention is as follows:

[0007] A first aspect of the present invention provides a method for constructing a preeclampsia prediction model based on temporal multimodal data, comprising the following steps:

[0008] Multimodal time-series feature data of pregnant women with known pregnancy outcomes were obtained to construct a time-series dataset; pregnant women with known pregnancy outcomes include those with preeclampsia and those without preeclampsia.

[0009] The graph attention network and the fully connected network were trained using the time series dataset to obtain a missing value fitting model;

[0010] The missing value fitting model is embedded as an embedding layer into a gated recurrent unit, and after training using the time series dataset, a preeclampsia prediction model is obtained.

[0011] Optionally, the steps for obtaining multimodal time-series feature data of pregnant women with known pregnancy outcomes and constructing a time-series dataset specifically include:

[0012] The pregnancy period of pregnant women with known pregnancy outcomes is divided into multiple time nodes. Multimodal feature data of pregnant women with known pregnancy outcomes are collected at each time node. The multimodal feature data of each time node are merged, and feature data that does not appear at the current time node is set to null. For repeated feature data, the last occurrence of the feature data is taken to obtain preprocessed multimodal time series feature data.

[0013] Using the frequency of feature data occurrences exceeding a preset value as a filtering criterion, the preprocessed multimodal time-series feature data is filtered to construct a time-series dataset.

[0014] Optionally, the step of constructing a time series dataset after filtering the preprocessed multimodal time series feature data based on the filtering criterion of the frequency of occurrence of feature data exceeding a preset value specifically includes:

[0015] Using the frequency of occurrence of feature data exceeding a preset value as a filtering criterion, the preprocessed multimodal time series feature data is filtered to obtain filtered multimodal time series feature data; the filtered multimodal time series feature data includes constant features and variable features;

[0016] In the filtered multimodal time series feature data, constant features are filled into the missing positions of constant features to obtain the filled multimodal time series feature data, and a time series dataset is constructed.

[0017] Optionally, the step of training the graph attention network and the fully connected network using the time series dataset to obtain the missing value fitting model specifically includes:

[0018] The time series dataset is converted into graph structure data, and the feature values ​​of a portion of the nodes in the graph structure data are masked as missing values ​​to form missing nodes, thus obtaining graph structure data with missing nodes.

[0019] After inputting the graph structure data with missing nodes into the graph attention network, the first data is output.

[0020] The first data is fused with the graph structure data to obtain fused feature data;

[0021] The fused feature data is expanded to obtain one-dimensional feature data;

[0022] The one-dimensional feature data is input into a three-layer stacked fully connected network to obtain the missing values ​​and the preeclampsia prediction results.

[0023] The first loss value is obtained based on the fill value and the actual value masked at the node, and the second loss value is obtained based on the preeclampsia prediction result and the actual preeclampsia result.

[0024] The parameters in the graph attention network and the fully connected network are adjusted based on the first loss value and the second loss value until the sum of the weights of the first loss value and the second loss value converges. Then, the trained graph attention network and the two stacked fully connected networks in the three stacked fully connected networks connected to the graph attention network constitute the missing value fitting model.

[0025] Optionally, after setting an information transmission mechanism in the graph attention network that allows missing nodes to only receive information from neighboring nodes and not transmit information to neighboring nodes, the graph structure data with missing nodes is input into the graph attention network, and the first data is output.

[0026] Optionally, the steps of embedding the missing value fitting model as an embedding layer into a gated recurrent unit and training it using the time series dataset to obtain a preeclampsia prediction model specifically include:

[0027] The missing value fitting model is embedded as an embedding layer into the input of a gated recurrent unit. The time series dataset is converted into graph structure data and used as input to obtain the probability of correctly predicting whether a pregnant woman has preeclampsia.

[0028] Based on the probability of correctly predicting whether a pregnant woman has preeclampsia and the actual situation of whether a pregnant woman has preeclampsia, the cross-entropy loss function is obtained;

[0029] The gated recurrent unit with the missing value fitting model is trained according to the cross-entropy loss function until the cross-entropy loss function converges, thus obtaining the preeclampsia prediction model.

[0030] In a second aspect, the present invention provides a method for predicting preeclampsia, wherein multimodal temporal feature data of a pregnant woman to be tested are input into a preeclampsia prediction model constructed using the construction method described above, thereby predicting whether the pregnant woman to be tested will develop preeclampsia in the future.

[0031] A third aspect of the present invention provides a preeclampsia prediction device, comprising:

[0032] The prediction unit is used to input the multimodal temporal feature data of the pregnant woman to be tested into the preeclampsia prediction model constructed using the construction method described above in this invention, and to predict whether the pregnant woman to be tested will develop preeclampsia in the future.

[0033] In a fourth aspect, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the prediction method of the present invention as described above.

[0034] A fifth aspect of the present invention provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, and when the computer program is executed by the processor, implements the prediction method of the present invention as described above.

[0035] Beneficial Effects: This invention constructs a missing value fitting model using graph attention networks and fully connected grids. This model effectively handles missing and irregular data in pregnant women's pregnancy testing data (such as medical indicators, vital signs, and HER data). It can extract effective features from highly missing data and utilize graph attention networks to capture potential correlations between various health data of pregnant women, minimizing the negative impact of missing values ​​on the model's prediction accuracy. Through the special processing of missing data by the missing value fitting model, missing information can be effectively recovered, ensuring the model can more accurately handle incomplete medical data. Simultaneously, by combining gated recurrent units for capturing and predicting temporal features, the accuracy of preeclampsia prediction is effectively improved. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating the construction method of a preeclampsia prediction model based on temporal multimodal analysis.

[0037] Figure 2 This is a statistical chart showing partial data and missing data.

[0038] Figure 3 This is a schematic diagram of constructing graph-structured data using features as nodes.

[0039] Figure 4 The diagram illustrates the information transmission adjustment strategy, where (a) is a schematic diagram of mutual information transmission between nodes, (b) is a schematic diagram of some nodes being replaced with vacant values, (c) is a schematic diagram of vacant nodes transmitting noise signals to adjacent nodes, and (d) is a schematic diagram of adjusting the information transmission strategy to reduce the impact of vacant values.

[0040] Figure 5 This is a framework diagram of a time-series multimodal preeclampsia prediction model.

[0041] Figure 6 This is a diagram of a time-series multimodal preeclampsia prediction model.

[0042] Figure 7 This is the confusion matrix for predicting preeclampsia.

[0043] Figure 8The effect of selecting different features on the predictive performance of the preeclampsia prediction model is shown in the figure. Detailed Implementation

[0044] This invention provides a method for constructing, predicting, and a predictive device for a preeclampsia prediction model based on temporal multimodal analysis. To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention is further described in detail below. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of the invention.

[0045] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0046] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of a described feature, integral, step, operation, element, and / or component, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" as used in this specification and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations. As used in this specification and the appended claims, the term "if" can be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrases "if determined" or "if [the described condition or event] is detected" can be interpreted, depending on the context, as meaning "once determined," "in response to determined," "once [the described condition or event] is detected," or "in response to the detection of [the described condition or event]." The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Many specific details are set forth in the following description to provide a thorough understanding of the present invention; however, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0047] This invention provides a method for constructing a preeclampsia prediction model based on temporal multimodal data, wherein, as... Figure 1 As shown, the construction method includes the following steps:

[0048] S1. Obtain multimodal time-series feature data of pregnant women with known pregnancy outcomes and construct a time-series dataset (containing multidimensional feature data); pregnant women with known pregnancy outcomes include those with preeclampsia and those without preeclampsia;

[0049] S2. Using the time series dataset, train the graph attention network (i.e., attention-driven graph convolutional network, or GAT for short) and the fully connected network to obtain the missing value fitting model;

[0050] S3 embeds the missing value fitting model as an embedding layer into a gated recurrent unit (i.e., a recurrent neural network unit with a gated mechanism, or GRU for short), and after training with the time series dataset, obtains a preeclampsia prediction model.

[0051] In practice, such as Figure 2 As shown, during fixed-time sampling, some data from pregnant women are often missing, with a missing rate as high as 90%, and it cannot be guaranteed that all tests can be performed at the fixed time intervals. Therefore, this invention constructs a missing value fitting model using GAT and a fully connected grid. This model effectively handles the missing and irregular data in pregnant women's pregnancy test data (such as medical indicators, vital signs, and HER data), extracts effective features from highly missing data, and uses GAT to capture the potential correlations between various health data of pregnant women, minimizing the negative impact of missing values ​​on the model's prediction accuracy. Through the special processing of missing data by the missing value fitting model, missing information can be effectively recovered, ensuring that the model can more accurately handle incomplete medical data. Simultaneously, by combining GRU for capturing and predicting temporal features, the prediction accuracy for preeclampsia is effectively improved.

[0052] Therefore, the preeclampsia prediction model constructed in this invention, by integrating the advantages of GAT and GRU, can effectively extract features even with missing data, extracting more effective features and fully exploring feature information that was previously unavailable due to missing data. When faced with highly missing HER data, it can effectively reduce the impact of missing data on prediction results, greatly enhancing the model's robustness in practical applications. Through multimodal data fusion and feature association analysis, this preeclampsia prediction model significantly improves the prediction accuracy of preeclampsia. The preeclampsia prediction model constructed in this invention can still stably predict preeclampsia even with severe data loss, a capability that is difficult to replace with other methods in actual clinical practice.

[0053] In this embodiment, the preeclampsia prediction model can identify potential preeclampsia risks in advance by analyzing multi-dimensional data of pregnant women during pregnancy; through time-series data of multiple time steps, the GRU network can capture the trend of changes in the health of pregnant women from historical data, and realize early warning of preeclampsia.

[0054] In addition, the experimental results show that the AUC (area under the ROC curve, i.e., the receiver operating characteristic curve) value of the time-series multimodal preeclampsia prediction model, namely the GAT-GRU model, for early prediction of preeclampsia reached 0.87, which has strong clinical application value.

[0055] In step S1, data for each pregnant woman is sampled periodically to ensure the temporal sequence of the data. This involves arranging all data records of the pregnant woman in chronological order from pregnancy to delivery and sampling at fixed time intervals. Data that is not detected at certain time points is marked as null, but the position of the time point is still retained, forming a continuous time series set.

[0056] In some implementations, the step of acquiring multimodal time-series feature data of pregnant women with known pregnancy outcomes and constructing a time-series dataset specifically includes steps S11 and S12:

[0057] S11. Divide the pregnancy period of pregnant women with known pregnancy outcomes (including those with preeclampsia and those without preeclampsia) into multiple time nodes, collect multimodal feature data of pregnant women with known pregnancy outcomes at each time node, merge the multimodal time series data of each time node, set the feature data that does not appear at the current time node to null, take the last appearance of the feature data that appears repeatedly, arrange the feature data obtained from multiple time nodes in chronological order from pregnancy to delivery, and obtain the preprocessed multimodal time series feature data.

[0058] The multimodal feature data of pregnant women at various time points exists in the form of text, numerical values, and categories. For example, electronic health records originate from various information systems in hospitals, containing various types and modalities of data, including records of various examinations performed on patients in the hospital. This data is composed of structured data (such as numerical data like blood pressure and blood sugar), semi-structured data (such as examination records and diagnostic conclusions), and text data (such as textual data in medical records), i.e., unstructured data. In other words, feature data is mainly divided into structured data, semi-structured data, and text data according to modality. Therefore, the multimodal feature data at each time point is merged. During the merging process, structured data is usually numerical data, such as blood pressure and blood sugar, which are already quantitative numerical features and can be used directly. Semi-structured data contains discrete categorical information (such as examination records and diagnostic conclusions), usually with a limited range of values. Specifically, categorical data is converted into sparse binary vectors. For example, the diagnostic result "positive / negative" can be converted into [1,0]. Text data (such as medical records) is unstructured data, rich in information but needs to be converted into numerical data through language processing. Specifically, a label encoder can be used to convert text data into numerical data.

[0059] The multimodal time-series feature data specifically includes medical indicators and vital signs. Medical indicators include outpatient visits, hospital admissions, diagnoses based on International Classification of Diseases and Related Health Problems (ICD-10) codes, and test results. Vital signs include body temperature, pulse, blood pressure, and heart rate. Some of these features include... Figure 2 As shown (where pain represents pain, weight_x represents weight at each sampling time point, nutrition represents nutrition, high risk score represents high risk score, high risk level represents high risk level, and fall indicates whether the pregnant woman has a history of falls).

[0060] Specifically, each gestational week (i.e., the week of pregnancy) in the pregnancy of pregnant women with known pregnancy outcomes can be used as a time node, i.e., a data collection node. Multimodal feature data of pregnant women with known pregnancy outcomes are collected in each gestational week. The multimodal feature data of each gestational week are merged, and feature data that does not appear in the current gestational week are set to null values. For repeated feature data, the last occurrence of the feature data is taken. The data are arranged in ascending order of gestational week (e.g., in the order of first week, second week, third week, fourth week, ...) to obtain preprocessed multimodal time series feature data.

[0061] More specifically, the 10th to 25th weeks of gestation for pregnant women with known pregnancy outcomes were divided into 16 time points, and multimodal feature data for each week of gestation from the 10th to the 25th week were used. This selection was made because this stage is a critical period where the risk of preeclampsia gradually emerges, which is helpful for early prediction. Furthermore, the data from this stage is generally more complete and standardized, which is beneficial for feature extraction in time series models. In addition, this selection method meets the model input requirements while including sufficient dynamic information and controlling the complexity of the time series length.

[0062] S12. Using the frequency of occurrence of feature data greater than a preset value as the filtering criterion, the preprocessed multimodal time series feature data is filtered to construct a time series dataset.

[0063] In some embodiments, the preprocessed multimodal time series feature data is filtered by using the frequency of occurrence of feature data exceeding a preset value as a filtering criterion to obtain filtered multimodal time series feature data. The filtered multimodal time series feature data includes constant features and variable features. The missing constant features in the filtered multimodal time series feature data are filled with constant features to obtain filled multimodal time series feature data, thus constructing a time series dataset.

[0064] Because many features in the preprocessed multimodal time-series feature data may be specific detections and not suitable for early screening of preeclampsia in routine pregnant women, a general screening process is performed based on the frequency and absence of feature data. Specifically, the screening criterion is that the frequency of feature data occurrences exceeds a preset value. After screening the preprocessed multimodal time-series feature data, those features with a frequency exceeding the preset value are selected to construct the time-series dataset. This screening strategy ensures that the model input data is more concise and effective, removing redundant and irrelevant features, thereby improving the accuracy and efficiency of prediction.

[0065] At each time point, the multimodal feature data for each pregnant woman includes constant features and variable features. Constant features are those that do not change over time and remain constant throughout the pregnancy. Variable features are those that change over time and evolve as the pregnancy progresses. For example, constant features include: blood type, age, family medical history (e.g., diabetes, hypertension), medical history (e.g., personal medical history, allergy history, surgical history, etc.), and fetal sex. Variable features include: blood pressure (e.g., systolic blood pressure, diastolic blood pressure, abnormal blood pressure), pulse, fundal height, fetal heart rate (e.g., abnormal fetal heart rate), blood parameters (e.g., white blood cells, red blood cells, platelets, mean platelet volume, hemoglobin concentration, etc.), biochemical indicators (e.g., triglycerides, urate, calcium, protein, etc.), and inflammatory markers (e.g., total white blood cell count, red blood cell distribution width, etc.). This invention particularly focuses on the changing trends of variable features because they are crucial in predicting preeclampsia and can reveal whether a pregnant woman has a potential risk of preeclampsia. For constant features, in this embodiment of the invention, the missing constant feature positions in the filtered multimodal time-series feature data are filled with constant features (specifically, the filling method can be: filling constant data with the mean or mode; for text data, using a label encoder to convert text data into numerical data), resulting in filled multimodal time-series feature data, which is then used to construct a time series dataset. That is, constant features are reused at each time point (for example, for the constant feature of blood type, when there are missing positions at certain time points, the blood type positions at these time points are filled with blood type data, thus this constant feature is reused at each time point), to help the model better extract the changing trends of variable features. In this way, the model can identify the dynamic changing patterns of variable features in the time series, thereby improving the early prediction ability of preeclampsia. This process ensures that the model not only focuses on the features of the current node but also effectively captures the long-term changing trends of the pregnant woman's health status.

[0066] In this step, the constructed time series dataset can be represented as: T = {T1, T2, ..., T} m}; where m is the total number of time points, and each T i This represents the set of all feature data at time point i, i.e., each time node T. i (i takes values ​​of 1, 2, ..., m) contains multidimensional feature data of the pregnant woman at that time point, which can be represented as: T i ={f i1 ,f i2 ,...f il}, where 1 ≤ l ≤ total characteristic number (e.g., 1 ≤ l ≤ 1048, 1 ≤ l ≤ 131, 1 ≤ l ≤ 95 or 1 ≤ l ≤ 31); fi1 f i2 f il The specific value or vectorized representation of a single feature at each time point.

[0067] In one specific embodiment, after step S11, 1048 features are obtained. Because of the uneven distribution of the various feature data during pregnancy testing, most of the 1048 features are specific detections and not suitable for screening preeclampsia in routine pregnant women. Therefore, the data is universally screened based on the frequency of occurrence and absence of a certain feature. Specifically, based on the criteria of a certain feature appearing more than 2000, 1500, 1200, and 1000 times in the overall data, 31, 79, 95, and 131 features are retained from the 1048 features, respectively, and time-series datasets are constructed using these four sets of features.

[0068] The 31 features include: systolic blood pressure, diastolic blood pressure, abnormal blood pressure, weight_x, abnormal weight, pulse, edema, fundal height, fetal heart rate, abnormal fetal heart rate, presenting part, abnormal fetal position, engagement, nutrition, fall, pain, functional needs, high-risk score, high-risk level, protein, urobilinogen, nitrite, hematuria, white blood cells, calcium oxalate, triphosphate, urate, hyaline casts, non-squamous epithelial cells, squamous epithelial cells, and mucus filaments.

[0069] The 79 features include: systolic blood pressure, diastolic blood pressure, abnormal blood pressure, weight_x, abnormal weight, pulse, edema, fundal height, fetal heart rate, abnormal fetal heart rate, presenting part, abnormal fetal position, engagement, nutrition, fall, pain, functional needs, high riskscore, high risk level, protein, urobilinogen, nitrite, hematuria, white blood cells, calcium oxalate, triphosphate, urate, hyaline casts, non-squamous epithelial cells, squamous epithelial cells, mucus filaments, as well as erythrocytes, fine granular casts, coarse granular casts, erythrocyte casts, white blood cell casts, vitamin C, plateletcrit, large platelet ratio, mean platelet volume, platelet distribution width, erythrocyte distribution width - CV (CV represents the coefficient of variation), erythrocyte distribution width - SD (SD represents the standard deviation), absolute basophil count, absolute eosinophil count, absolute neutrophil count, absolute monocyte count, and lymphocytes. Absolute values, percentage of basophils, percentage of eosinophils, percentage of neutrophils, percentage of monocytes, percentage of lymphocytes, platelets, mean corpuscular hemoglobin concentration, mean corpuscular hemoglobin content, mean corpuscular volume, hematocrit, hemoglobin, atypical lymphocytes, total white blood cell count, absolute value of nucleated red blood cells, percentage of nucleated red blood cells, glucose, pH, indirect bilirubin, total protein, albumin, total bilirubin, direct bilirubin, gamma-glutamyl transferase, total bile acids, creatinine, uric acid, A / G (serum albumin to globulin ratio), globulin, cystatin C, bilirubin, and ketone bodies.

[0070] The 95 characteristics include: systolic blood pressure, diastolic blood pressure, abnormal blood pressure, weight_x, abnormal weight, pulse, edema, fundal height, fetal heart rate, abnormal fetal heart rate, presenting part, abnormal fetal position, engagement, nutrition, fall, pain, functional needs, high riskscore, and high risk. Levels of blood, protein, urobilinogen, nitrite, hematuria, white blood cells, calcium oxalate, triphosphate, urate, hyaline casts, non-squamous epithelial cells, squamous epithelial cells, mucus filaments, as well as erythrocytes, fine granular casts, coarse granular casts, erythrocyte casts, white blood cell casts, vitamin C, plateletcrit, large platelet ratio, mean platelet volume, platelet distribution width, erythrocyte distribution width-CV, erythrocyte distribution width-SD, absolute basophil count, absolute eosinophil count, absolute neutrophil count, absolute monocyte count, absolute lymphocyte count, percentage of basophils, percentage of eosinophils, percentage of neutrophils, percentage of monocytes, percentage of lymphocytes, platelets, mean corpuscular hemoglobin concentration. Mean corpuscular hemoglobin (MCH), mean corpuscular volume (MCV), hematocrit, hemoglobin, atypical lymphocytes, total white blood cell count, absolute nucleated red blood cell count, percentage of nucleated red blood cells, glucose, pH, indirect bilirubin, total protein, albumin, total bilirubin, direct bilirubin, gamma-glutamyl transferase (GGT), total bile acids, creatinine, uric acid, A / G ratio, globulin, cystatin C, bilirubin and ketone bodies, as well as specific gravity, alanine aminotransferase (ALT), aspartate aminotransferase (AST), spouse's health, height, weight_y (weight at the time of the first prenatal checkup), BMI (body mass index), hypertension, parity, fetal movement at gestational age, early pregnancy symptoms, weeks of early pregnancy symptoms, diabetes, no history of disease, no history of drug allergies, and last menstrual period.

[0071] The 131 features include: systolic blood pressure, diastolic blood pressure, abnormal blood pressure, weight_x, abnormal weight, pulse, edema, fundal height, fetal heart rate, abnormal fetal heart rate, presenting part, abnormal fetal position, engagement, nutrition, fall, pain, functional needs, high riskscore, and high risk. Levels of blood, protein, urobilinogen, nitrite, hematuria, white blood cells, calcium oxalate, triphosphate, urate, hyaline casts, non-squamous epithelial cells, squamous epithelial cells, mucus filaments, and red blood cells, fine granular casts, coarse granular casts, red blood cell casts, white blood cell casts, vitamin C, plateletcrit, large platelet ratio, mean platelet volume, platelet distribution width, red blood cell distribution width (CV), red blood cell distribution width (SD), absolute basophil count, absolute eosinophil count, absolute neutrophil count, absolute monocyte count, absolute lymphocyte count, percentage of basophils, percentage of eosinophils, percentage of neutrophils, percentage of monocytes, percentage of lymphocytes, platelets, mean corpuscular hemoglobin concentration, mean corpuscular hemoglobin content, mean corpuscular volume, hematocrit, hemoglobin, atypical lymphocytes, total white blood cell count, absolute count of nucleated red blood cells, percentage of nucleated red blood cells, glucose. pH, indirect bilirubin, total protein, albumin, total bilirubin, direct bilirubin, gamma-glutamyl transferase, total bile acids, creatinine, uric acid, A / G ratio, globulin, cystatin C, bilirubin and ketone bodies, as well as specific gravity, alanine aminotransferase, aspartate aminotransferase, spouse's health, height, weight (y), BMI, hypertension, parity, fetal movement, gestational age, early pregnancy reactions, early pregnancy reaction weeks, diabetes, no history of disease, no medication. History of drug allergies, last menstrual period, and levels of calcium, alkaline phosphatase, potassium, sodium, magnesium, atypical lymphocytes, chloride, glycated hemoglobin (HPLC), lactate dehydrogenase, creatine kinase, α-hydroxybutyrate dehydrogenase, high-sensitivity CRP, estradiol, white blood cell morphology, platelet morphology, α-amylase, red blood cell morphology, lipase, cholinesterase, creatine kinase isoenzyme, risk value, free hCGβ, gestational age at sampling, number of fetuses, expected age at delivery, gestational age at ultrasound, and PAPP-A. MoM value (PAPP-A refers to pregnancy-associated plasma protein A, and the PAPP-A MoM value is an important value in the prenatal screening test for Down syndrome), nuchal translucency thickness (direct measurement value), NT MoM value (NT value refers to the thickness of the fetal nuchal translucency measured under B-ultrasound, and the NT MoM value is a standardized multiple of the nuchal translucency thickness), free T4, thyroid-stimulating acid, rapid C-reactive protein (instrument), reticulocyte percentage, immature reticulocyte percentage, neutrophil band nucleus, and triglycerides.

[0072] In step S2, such as Figure 5As shown, in some embodiments, the steps of training the graph attention network and the fully connected network using the time-series dataset specifically include steps S21 to S27:

[0073] S21. Convert the time series dataset into graph structure data, and mask the feature values ​​of some nodes in the graph structure data as missing values ​​to form missing nodes, thus obtaining graph structure data with missing nodes.

[0074] Specifically, at the input end, graph-structured data is constructed using features as nodes (in some specific examples, such as...). Figure 3 As shown, a graph structure is constructed using features of a gestational week as nodes. Converting a time-series dataset into graph structure data means that in this data, each node represents a feature, and each feature value is called an attribute of that node. Typically, the expression for a graph is:

[0075] G = (V, E)

[0076] Where V is the total set of node features in the graph, and E is the edge set. Node features V = {v1, v2, v3, ..., v...} m}; Each node v i ∈V has multiple feature attributes, which serve as nodes v i The attribute value, v i In this context, i takes the values ​​1, 2, 3, ..., m. Every two nodes (i.e., v...) i and v j (v) are connected by an edge (v) i ,v j The nodes are connected, and the edge features together form an edge set E. The weight of each edge represents the correlation between the nodes. In this way, time series data can be transformed into a graph structure, which effectively represents the correlation between each feature and the contextual information of the time series. This graph structure provides a suitable input format for subsequent prediction models and helps to capture the complex relationships between features.

[0077] S22, such as Figure 5 As shown, after inputting the graph structure data with missing nodes into GAT, the first data is output.

[0078] In some specific embodiments, such as Figure 4 and Figure 5 As shown, after setting an information transmission mechanism in the graph attention network that allows missing nodes to only receive information from neighboring nodes and not transmit information to neighboring nodes, the graph structure data with missing nodes is input into the graph attention network, and the first data is output.

[0079] More specifically, a propagation threshold is set in the GAT network structure. For example, nodes with missing node values ​​less than 0 are not allowed to propagate feature data to neighboring nodes. This prevents these nodes from passing features to neighboring nodes and maximizes the transmission of feature information from neighboring nodes to missing nodes. The results show that setting a higher threshold makes it easier for missing nodes to obtain information from neighboring nodes, enhances the model's ability to fit and fill missing data, and improves the prediction performance of subsequent time series models.

[0080] In some specific embodiments, GAT has a three-layer structure.

[0081] The first data output after GAT processing (corresponding to) Figure 5 The output (01) on the left and the graph structure data of the original input (corresponding) Figure 5 If the lengths of the inputs on the left are consistent after one-dimensional expansion, then the first data can be represented as: Output features (i = 1, 2, ..., N) represents the final node representation of the U-th layer. At this point, the relationships between nodes have been optimized. Through the graph attention mechanism, the model can more accurately capture the potential connections between nodes, thereby improving the accuracy of feature extraction and the model's predictive performance. This step provides high-quality feature input for subsequent preeclampsia prediction.

[0082] In steps S21 and S22, to simulate the common missing data situations in actual medical data, some nodes in the graph structure data are intentionally masked, and the feature values ​​of some nodes are replaced with missing values. Simultaneously, for these missing nodes, an information transmission mechanism is specifically designed so that these missing nodes can only receive information from neighboring nodes and will not transmit information to neighboring nodes. Specifically, as follows... Figure 4 As shown in (d) in the diagram. This method reduces the interference of missing nodes on the information transmission between adjacent nodes, ensuring that the model is not negatively affected by missing data during training. Specifically, by adjusting the propagation threshold of missing nodes (for example, setting no feature transmission when the feature value of a missing node is less than 0; specifically, during training, the dataset is randomly divided into training and test sets in an 8:2 ratio, and the training set is trained 96 times, each time setting one node value (an entire column) to -9, and adding constraints to limit the external transmission of information by nodes with values ​​less than -9), the model can effectively control the noise caused by missing data and maximize the transmission of effective information between adjacent nodes, further improving the prediction performance of the time series model. This invention artificially sets the feature values ​​of some nodes to be missing and designs a specific propagation mechanism so that missing nodes only receive feature information from adjacent nodes and do not transmit feature information outwards. This design effectively reduces the interference of missing nodes on information transmission during model training and improves the model's ability to fit and fill in missing data.

[0083] Graph structure data with missing values ​​is input into a three-layer graph neural network based on a graph attention mechanism, namely a three-layer GAT. The graph structure data input to the GAT includes node features and graph structure information, with each node v... i The features are represented as Where x ij Let v be the j-th feature of the i-th node. Graph structure information is represented by the edge set E, describing the connections between nodes. In GAT, node features are updated through an attention mechanism. At layer u, node v... i The feature update formula can be expressed as:

[0084] 1≤u≤3;

[0085] in, Represents node v i The feature vector at the (u+1)th layer, Represents node v j The feature vector at the u-th layer, This represents the attention coefficient of the u-th layer, i.e., node v. i Its neighboring node v j The correlation strength represents the weighting coefficient in the graph attention mechanism, W. (u) b is the weight matrix of the u-th layer. (u) σ is the bias term, σ is the activation function, and N(v) is the bias term. i ) refers to the relationship with node v i The set of directly connected nodes.

[0086] This network can adaptively assign different weights based on the relationships between nodes, thereby effectively extracting the potential correlations between nodes in the graph. When dealing with missing nodes, it can automatically adjust the way node information is propagated, minimizing the information interference caused by missing nodes. This mechanism can improve the model's learning ability under conditions of missing data.

[0087] S23, such as Figure 5 As shown, the first data (i.e., H) (U) ,correspond Figure 5 The output 01 on the left and the graph structure data (denoted as G, corresponding to) Figure 5 The input on the left is fused to obtain fused feature data (T', corresponding to...). Figure 5 The input on the left is 02); then T' = G + H (U) .

[0088] In some embodiments, the fusion method is residual connection. Residual connection is a structure used in deep neural networks to improve the training performance of deep networks. Its core idea is to introduce "shortcuts" or "skip connections" into the network, allowing the input to be directly passed to subsequent layers and added to the output after one or more layers of processing. This structure allows the network to learn an identity mapping of the input, thereby alleviating the gradient vanishing and gradient exploding problems in deep network training. Therefore, in this embodiment, using residual connection as a fusion method helps to preserve key information in the original data and combine the features extracted by the graph neural network with the original features.

[0089] S24. Expand the fused feature data to obtain one-dimensional feature data (i.e., T). flat ), that is, T flat =Flatten(T′); Flatten indicates a one-dimensional expansion.

[0090] S25, such as Figure 5 As shown, the one-dimensional feature data is input into a three-layer stacked fully connected network to obtain the missing values ​​and the preeclampsia prediction results;

[0091] S26. Obtain the first loss value based on the fill value and the actual value of the masked node, and obtain the second loss value based on the preeclampsia prediction result and the actual preeclampsia result.

[0092] S27. Adjust the parameters in the graph attention network and the fully connected network according to the first loss value and the second loss value until the weights of the first loss value and the second loss value converge. Then, the trained graph attention network and the two stacked fully connected networks in the three stacked fully connected networks connected to the graph attention network constitute the missing value fitting model.

[0093] In this step, such as Figure 5 As shown, the three-layer stacked fully connected network includes a first layer of fully connected network (corresponding to...). Figure 5 The fourth layer), the second layer fully connected network (corresponding to) Figure 5 Layer 5) and Layer 3 fully connected network (corresponding to Figure 5 The sixth layer of the GAT (Gate Amplifier) ​​is used to input one-dimensional feature data into the first fully connected network. The output of the first fully connected network is used as the input of the second fully connected network, and the output of the second fully connected network is used as the input of the third fully connected network. The trained three-layer GAT and the first and second fully connected networks connected to the three-layer GAT in sequence constitute the missing value fitting model.

[0094] In one specific embodiment, during masking, the feature values ​​of a portion of the nodes in the graph structure data are masked (denoted as mask01, used to set the mask for missing values, which is used to calculate the loss of the missing value filling task) as missing values ​​to form missing nodes, resulting in graph structure data 01 with missing nodes; the feature values ​​of another portion of the nodes in the graph structure data are masked (denoted as mask02, for the preeclampsia classification task, used to calculate the loss of the preeclampsia classification task) as missing values ​​to form missing nodes, resulting in graph structure data 02 with missing nodes.

[0095] The graph structure data 01 with missing nodes is input into a graph attention network (with an information transmission mechanism that allows missing nodes to only receive information from neighboring nodes and not pass information to neighboring nodes), and the first data is output. The first data is fused with the graph structure data to obtain fused feature data. The fused feature data is expanded to obtain one-dimensional feature data. The one-dimensional feature data is input into a three-layer stacked fully connected network to obtain the filling value for the missing values ​​(the model will automatically fill a filling value at the feature position of the node with the mask). The first loss value is obtained based on the filling value and the true value of the masked node.

[0096] Next, the graph structure data 02 with missing nodes is input into a graph attention network (with an information transmission mechanism that allows missing nodes to only receive information from neighboring nodes and not transmit information to neighboring nodes), and the first data is output. The first data is then fused with the graph structure data to obtain fused feature data. The fused feature data is then expanded to obtain one-dimensional feature data. The one-dimensional feature data is then input into a three-layer stacked fully connected network to obtain the preeclampsia prediction result (i.e., the model's prediction output for preeclampsia classification). The second loss value is obtained based on the preeclampsia prediction result and the actual preeclampsia result (i.e., the actual label of preeclampsia).

[0097] The parameters in the graph attention network and the fully connected network are adjusted based on the first loss value and the second loss value until the sum of the weights of the first loss value and the second loss value converges.

[0098] The weighted sum of the first and second loss values ​​equals the loss value. GAT loss GAT= α × loss01 + β × loss02, where α = 0.8 and β = 0.2; the first loss value, loss01, measures the model's ability to fill in missing values ​​(i.e., its ability to fit missing values), and the second loss value, loss02, measures the model's ability to extract features related to preeclampsia. Because the main function of the missing value fitting model is its ability to fill in missing values ​​(or its fitting ability), loss01 has a larger weight. This loss value... GAT Training the model until the loss value converges allows it to effectively handle missing data while ensuring its accuracy in predicting the risk of preeclampsia.

[0099] In this step, the one-dimensional feature data T flat The input is fed into a three-layer stacked fully connected network, which will eventually output a binary classification result y (then mask02 and loss02 mentioned above are related to this target output y).

[0100] y=σ(W3·σ(W2·σ(W1·T flat +b1)+b2)+b3)

[0101] Where W1, W2, and W3 are the weight matrices of each layer, b1, b2, and b3 are the biases of each layer, and σ is the activation function.

[0102] Fully connected networks extract more abstract features through non-linear mapping of data. The design of fully connected layers can improve the model's ability to represent features and enhance prediction accuracy.

[0103] In step S3, such as Figure 5 As shown, in some embodiments, the steps of embedding the missing value fitting model as an embedding layer into a gated recurrent unit and training it using the time series dataset to obtain a preeclampsia prediction model specifically include:

[0104] S31. The missing value fitting model is embedded as an embedding layer into the input of the gated recurrent unit. The time series dataset is converted into graph structure data and used as input to obtain the probability of correctly predicting whether a pregnant woman has preeclampsia.

[0105] S32. Based on the probability of correctly predicting whether a pregnant woman has preeclampsia and the actual situation of whether a pregnant woman has preeclampsia, the cross-entropy loss function is obtained.

[0106] Specifically, the cross-entropy loss function is:

[0107]

[0108] in, Let be the total loss function, representing the prediction error on the current model dataset; N is the total number of samples; y n This indicates whether the nth sample has preeclampsia. Its value is 0 or 1, where 0 means that the sample does not have preeclampsia and 1 means that the sample has preeclampsia. represents the predicted probability of the nth sample (i.e., the probability of correctly predicting whether a pregnant woman has preeclampsia), and represents the confidence of the model in predicting that the sample belongs to category 1 (i.e., the category of having preeclampsia in the target classification task), ranging from 0 to 1.

[0109] S33. Train the gated recurrent unit with the missing value fitting model according to the cross-entropy loss function (the dataset can be divided into 80% training set and 20% test set during training) until the cross-entropy loss function converges to obtain the preeclampsia prediction model.

[0110] Based on the above construction, the resulting preeclampsia prediction model has the following confusion matrix for preeclampsia prediction: Figure 7 As shown, the performance metrics of this preeclampsia prediction model are AUC ≥ 0.87, sensitivity ≥ 76%, specificity ≥ 94%, and F1 score (the harmonic mean of precision and recall) of 0.85. An AUC ≥ 0.87 indicates that the model has good overall classification ability for predicting preeclampsia; a sensitivity ≥ 76% indicates excellent sensitivity in identifying preeclampsia; a specificity ≥ 94% indicates high specificity in excluding non-preeclampsia patients; and an F1 score of 0.85 indicates good balance between precision and sensitivity. Because the dataset is imbalanced, although the accuracy is high, this invention primarily uses the AUC value as the main evaluation metric.

[0111] In this step, the principle of the gated loop unit is as follows: if the input data is The output of each layer serves as the input to the next layer, forming a recurrent neural network used to process time series data. The network update formula is:

[0112]

[0113] Among them, h m Let z be the hidden state at time step m. m It's an update gate. These are candidate hidden states. Each layer of the network contains 256 units, effectively capturing contextual information in time series data.

[0114] In this invention, a pre-trained missing value fitting model is used as an embedding layer to extract features from the input data and fit missing values. After the preeclampsia prediction model is trained, assuming the acquired time-series data of the pregnant woman to be tested is T... seq ={T″1,T″2,...,T″ m}, where T″ i Let represent the feature vector at time step i (i takes the values ​​1, 2, ..., m). Then, the data at the input (e.g., ...) is... Figure 5 The input on the right side (as shown) is transformed into graph-structured data. The embedding layer uses GAT to extract features from each time series node using the graph-structured data, capturing the interrelationships between nodes, and providing more accurate input features for subsequent prediction models.

[0115]

[0116] Among them, F(T″ i ) indicates time step T i The result of embedding the features is denoted as . The graph structure features of the nodes are extracted using GAT. It is a mapping function of the embedding layer. i represents the index of the time step. Since the entire time series contains m time steps, i = 1, 2, ..., m, indicating that the i-th time node in the sequence is being processed.

[0117] In this embodiment, the Adam optimizer can be used to further optimize the model during training, with the initial learning rate set to 0.001. These settings effectively avoid overfitting and accelerate model convergence, enabling the model to learn effective feature representations in a shorter time and improving prediction accuracy.

[0118] like Figure 6 As shown, the multimodal temporal feature data of the pregnant woman to be tested is transformed into graph structure data at the input end and then input into the trained preeclampsia prediction model. The input data is processed by the embedding layer for feature extraction and missing value fitting. Then, the GRU is used to perform the temporal feature step and prediction. Finally, the output of each time step is obtained through the linear layer, and the output of the last time step is selected as the final binary classification result.

[0119] In the above text, 31, 79, 95, and 131 features were selected respectively to construct time series datasets. For example... Figure 8As shown, the model achieves the best classification prediction performance when 79 features are selected. This means that using 79 features allows for the acquisition and preservation of relationships between features, enabling the model to identify more effective features even with missing values, and reducing the impact of missing values. The data used in this invention includes pregnant women's test results, family medical history, disease history, physical signs, and health factors during pregnancy as features. Based on modality, the data is mainly divided into numerical and textual types. The features are selected based on their missing value after time-series sampling, choosing those with low missing rates. Each feature is considered a node, and the correlation coefficients between features are used as edges to construct a graph structure. Each node in the constructed graph structure is treated as having the same dimension.

[0120] In addition, this invention also uses the GCN (Graph Convolutional Network) model to process the gaps. The overall prediction effect of the model is poor, the comparison results are not obvious, and different filling methods can be barely obtained. This has a certain impact on the fitting ability of GCN. Furthermore, whether the feature data is normalized has the opposite effect on the same filling method.

[0121] Table 1. Impact of different processing methods on missing nodes

[0122]

[0123] This invention also provides a method for predicting preeclampsia, wherein the multimodal temporal feature data of the pregnant woman to be tested are input into the preeclampsia prediction model constructed using the construction method described above in this invention, and the method is used to predict whether the pregnant woman to be tested will develop preeclampsia in the future.

[0124] This invention also provides a preeclampsia prediction device, comprising:

[0125] The prediction unit is used to input the multimodal temporal feature data of the pregnant woman to be tested into the preeclampsia prediction model constructed by the construction method described above in the embodiments of the present invention, and to predict whether the pregnant woman to be tested will develop preeclampsia in the future.

[0126] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the protein structure model training method described above or the protein structure prediction method described above.

[0127] The computer-readable medium described in this embodiment may be a computer-readable storage medium, a computer-readable signal medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to, an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), or any suitable combination thereof.

[0128] The present invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program that can run on the processor. When the computer program is executed by the processor, it implements the protein structure model training method described above in the embodiments of the present invention or the protein structure prediction method described above in the embodiments of the present invention.

[0129] In this embodiment, the memory can be volatile memory, such as random access memory; the memory can also be non-volatile memory, such as read-only memory, flash memory, hard disk, etc. The processor can be a central processing unit, controller, microcontroller, microprocessor, or other data processing chip.

[0130] In summary, this invention provides a method for constructing, predicting, and using a temporal multimodal preeclampsia prediction model. This preeclampsia prediction model effectively addresses the shortcomings of existing technologies in multimodal data processing and high missing data issues. Unlike traditional preeclampsia prediction methods, this invention utilizes GAT (Generative Attraction-Based Analysis) for feature extraction, particularly effective in extracting features from highly missing data, especially in cases of high missing data rates. Combining this with a gated recurrent unit (GRU) for temporal feature capture and prediction significantly improves the accuracy of preeclampsia prediction. Traditional preeclampsia prediction methods typically rely on single clinical indicators and are often used for later stages of the disease. This invention, however, analyzes multidimensional data from the pregnant woman's pregnancy to identify potential preeclampsia risks in advance. Through multi-timestep temporal data input, the GRU network can capture trends in the pregnant woman's health from historical data, enabling early warning of preeclampsia. In conclusion, this invention proposes a novel strategy for handling highly missing data by combining GAT feature extraction technology with GRU temporal prediction. Missing values ​​are a common problem in EHR data, and traditional methods cannot effectively address this challenge. The predictive model of this invention can still effectively extract features even with a large number of missing values, greatly enhancing its robustness in practical applications. The predictive model provided by this invention not only improves the accuracy of preeclampsia prediction but also offers a more efficient data processing and prediction tool for clinical practice.

[0131] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for constructing a preeclampsia prediction model based on temporal multimodal data, characterized in that, Includes the following steps: Multimodal time-series feature data of pregnant women with known pregnancy outcomes were obtained to construct a time-series dataset; pregnant women with known pregnancy outcomes include those with preeclampsia and those without preeclampsia. The graph attention network and the fully connected network were trained using the time series dataset to obtain a missing value fitting model; The missing value fitting model is embedded as an embedding layer into a gated recurrent unit, and after training with the time series dataset, a preeclampsia prediction model is obtained. The specific steps for obtaining multimodal time-series feature data of pregnant women with known pregnancy outcomes and constructing a time-series dataset include: The pregnancy period of pregnant women with known pregnancy outcomes is divided into multiple time nodes. Multimodal feature data of pregnant women with known pregnancy outcomes are collected at each time node. The multimodal feature data of each time node are merged, and feature data that does not appear at the current time node is set to null. For repeated feature data, the last occurrence of the feature data is taken to obtain preprocessed multimodal time series feature data. Using the frequency of feature data occurrences exceeding a preset value as a filtering criterion, the preprocessed multimodal time-series feature data is filtered to construct a time-series dataset; The steps for constructing a time series dataset after filtering the preprocessed multimodal time series feature data based on the filtering criterion of the frequency of occurrence of feature data exceeding a preset value specifically include: Using the frequency of occurrence of feature data exceeding a preset value as a filtering criterion, the preprocessed multimodal time series feature data is filtered to obtain filtered multimodal time series feature data; the filtered multimodal time series feature data includes constant features and variable features; In the filtered multimodal time series feature data, constant features are filled into the missing positions of constant features to obtain the filled multimodal time series feature data, and a time series dataset is constructed. The specific steps for training the graph attention network and the fully connected network using the time series dataset to obtain the missing value fitting model include: The time series dataset is converted into graph structure data, and the feature values ​​of a portion of the nodes in the graph structure data are masked as missing values ​​to form missing nodes, thus obtaining graph structure data with missing nodes. After inputting the graph structure data with missing nodes into the graph attention network, the first data is output. The first data is fused with the graph structure data to obtain fused feature data; The fused feature data is expanded to obtain one-dimensional feature data; The one-dimensional feature data is input into a three-layer stacked fully connected network to obtain the missing values ​​and the preeclampsia prediction results. The first loss value is obtained based on the fill value and the actual value masked at the node, and the second loss value is obtained based on the preeclampsia prediction result and the actual preeclampsia result. The parameters in the graph attention network and the fully connected network are adjusted based on the first loss value and the second loss value until the sum of the weights of the first loss value and the second loss value converges. Then, the trained graph attention network and the two stacked fully connected networks in the three stacked fully connected networks connected to the graph attention network constitute the missing value fitting model.

2. The construction method according to claim 1, characterized in that, After setting up an information transmission mechanism in the graph attention network that allows missing nodes to only receive information from neighboring nodes and not pass information to neighboring nodes, the graph structure data with missing nodes is input into the graph attention network, and the first data is output.

3. The construction method according to claim 2, characterized in that, The steps of embedding the missing value fitting model as an embedding layer into a gated recurrent unit and training it using the time series dataset to obtain the preeclampsia prediction model specifically include: The missing value fitting model is embedded as an embedding layer into the input of a gated recurrent unit. The time series dataset is converted into graph structure data and used as input to obtain the probability of correctly predicting whether a pregnant woman has preeclampsia. Based on the probability of correctly predicting whether a pregnant woman has preeclampsia and the actual situation of whether a pregnant woman has preeclampsia, the cross-entropy loss function is obtained; The gated recurrent unit with the missing value fitting model is trained according to the cross-entropy loss function until the cross-entropy loss function converges, thus obtaining the preeclampsia prediction model.

4. A method for predicting preeclampsia, characterized in that, The multimodal temporal feature data of the pregnant woman to be tested are input into the preeclampsia prediction model constructed using the construction method described in any one of claims 1-3, so as to predict whether the pregnant woman to be tested will develop preeclampsia in the future.

5. A device for predicting preeclampsia, characterized in that, include: The prediction unit is used to input the multimodal temporal feature data of the pregnant woman to be tested into the preeclampsia prediction model constructed using the construction method described in any one of claims 1-3, and to predict whether the pregnant woman to be tested will develop preeclampsia in the future.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the prediction method of claim 4.

7. An electronic device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and when the computer program is executed by the processor, it implements the prediction method of claim 4.

Images