Periconceptional birth defect risk assessment method, device and electronic equipment

Abstract

The application provides a periconceptional birth defect risk assessment method and device and electronic equipment, and relates to the technical field of medical information. The method comprises the following steps: in response to individual multi-dimensional data input by a target object, the individual multi-dimensional data is mapped into individual variable information, and observed variables and missing variables are determined; based on the observed variables and a preset causal network, the missing variables are inferred posteriorly to obtain a posterior distribution; based on the observed variables and the posterior distribution, fact state risk information is generated; after determining an intervention variable, counterfactual intervention is performed on the intervention variable, and based on the observed variables and the posterior distribution, counterfactual state risk information of the intervention variable is generated; and based on the preset causal network, the fact state risk information and the counterfactual state risk information of the intervention variable are used to generate a risk assessment result. The application introduces counterfactual intervention to quantitatively predict the intervention effect, and simultaneously solves the problems of complex dependence and data loss through the causal network, so that the assessment result is more reliable.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a method, device, and electronic device for assessing the risk of birth defects during the perinatal period. Background Technology

[0002] Birth defects are a leading cause of infant death and disability, placing a heavy burden on families and society. Traditional perinatal birth defect risk assessment relies primarily on qualitative assessments using epidemiological risk factor lists. This approach struggles to quantify individual risk and effectively handles complex interactions among multiple factors. To overcome these limitations, existing technologies have developed risk assessment tools based on parametric statistical models, such as logistic regression models. However, these models typically assume independence between risk factor variables, making it difficult to effectively model the complex, non-linear dependencies that exist between risk factors in the real world. Furthermore, these models have limited capabilities when dealing with incomplete individual data, reducing the accuracy and practicality of the assessment. Simultaneously, there have been attempts to use probabilistic graphical models for medical diagnosis. These methods can perform probabilistic inference based on evidence input from the target population and handle missing data relatively well. However, these methods typically only identify the degree of risk and cannot provide information on risk changes after evidence modification. In other words, existing technologies primarily remain at the risk assessment level, lacking the ability to quantitatively simulate and predict the effects of interventions. Summary of the Invention

[0003] In order to overcome the shortcomings of the prior art, the present invention provides a method, device and electronic device for assessing the risk of birth defects during the perinatal period.

[0004] According to a first aspect of the present invention, a method for assessing the risk of birth defects during the perinatal period is provided, comprising: responding to individual multidimensional data input by a target subject, mapping the individual multidimensional data to individual variable information corresponding to a preset causal network, and determining observed variables and missing variables; based on the observed variables and the preset causal network, performing posterior inference on the missing variables to obtain a posterior distribution corresponding to the missing variables; based on the observed variables and the posterior distribution, inferring the risk value of the target outcome and generating factual state risk information; after determining the intervention variable according to the individual variable information and a preset set of interventionable variables, performing counterfactual intervention on the intervention variable, and inferring the risk value of the target outcome under the intervention state based on the observed variables and the posterior distribution, generating counterfactual state risk information of the intervention variable; and based on the preset causal network, generating a risk assessment result of the intervention variable according to the factual state risk information and the counterfactual state risk information of the intervention variable.

[0005] Optionally, the pre-constructed causal network is constructed in advance as follows: a set of candidate risk variables is determined based on literature evidence, historical sample data, and expert knowledge; a constrained structural learning algorithm is used to test the conditional independence of the candidate risk variable set to generate a network skeleton; temporal constraints are established according to the order of occurrence of variables at different stages of the perinatal period; edge constraints are established based on temporal constraints and expert knowledge; edge constraints include prohibited edge constraints and / or mandatory edge constraints; the network structure is optimized using a scoring search algorithm based on the network skeleton, temporal constraints, and edge constraints to generate a target directed acyclic graph; training samples are constructed based on historical sample data, and parameter learning is performed on the target directed acyclic graph based on the training samples to obtain the conditional probability parameters corresponding to each node; the pre-constructed causal network is constructed based on the target directed acyclic graph and the conditional probability parameters.

[0006] Optionally, in response to the individual multidimensional data input by the target object, the individual multidimensional data is mapped to individual variable information corresponding to a preset causal network, and the observed variables and missing variables are determined. This includes: in response to the individual multidimensional data input by the target object, the individual multidimensional data is validated to obtain processed individual data; the individual multidimensional data includes individual data corresponding to at least one stage of the target object in preconception, early pregnancy, mid-pregnancy, and / or late pregnancy; based on a preset variable dictionary, each processed individual data is mapped to variable information of the corresponding node in the preset causal network to obtain a set of variable information that corresponds one-to-one with the individual multidimensional data; based on predefined conditions in the preset causal network, variables in the variable information set that meet the predefined conditions are determined as observed variables, and variables in the variable information set that do not meet the predefined conditions are determined as missing variables; the predefined conditions include at least one of integrity verification conditions, legal value range conditions, logical consistency conditions, and invalid identifier filtering conditions.

[0007] Optionally, based on the observed variables and a pre-defined causal network, posterior inference is performed on the missing variables to obtain the posterior distribution corresponding to the missing variables, including: determining the candidate state range of the missing variables according to the relevant nodes of the observed variables in the pre-defined causal network; the relevant nodes include at least neighboring nodes, ancestor nodes, and temporally preceding nodes; performing posterior inference on the missing variables to obtain the posterior distribution; the posterior distribution includes a univariate posterior distribution and / or a joint posterior distribution.

[0008] Optionally, based on the observed variables and the posterior distribution, the risk value of the target outcome is inferred to generate factual state risk information, including: according to the observed variables, multiple candidate states in the posterior distribution are weighted according to the posterior probability to determine the risk value of the target outcome in each candidate state; the risk values ​​in each candidate state are weighted and summarized to obtain the risk value of the target outcome and generate factual state risk information.

[0009] Optionally, after determining the intervention variable based on individual variable information and a pre-defined set of interventionable variables, counterfactual intervention is performed on the intervention variable. Based on the observed variables and the posterior distribution, the risk value of the target outcome under the intervention state is inferred, generating counterfactual state risk information of the intervention variable. This includes: determining at least one intervention variable from the pre-defined set of interventionable variables based on individual variable information; setting the intervention variable as the target intervention state according to pre-defined intervention rules, and suppressing the influence of the original causal input of the intervention variable on its state in a pre-defined causal network; calculating the risk value of the target outcome under each candidate state by weighting multiple candidate states according to the posterior probability based on the observed variables and the intervention variable; and summarizing the risk values ​​under each candidate state by weight to obtain the risk value of the target outcome, generating counterfactual state risk information of the intervention variable.

[0010] Optionally, based on a pre-defined causal network, and according to the factual state risk information and the counterfactual state risk information of the intervention variable, a risk assessment result for the intervention variable is generated, including: generating counterfactual effect information of the intervention variable based on the factual state risk information and the counterfactual state risk information of the intervention variable; the counterfactual effect information of the intervention variable includes the risk difference, risk ratio, and / or risk change magnitude of the target outcome before and after the intervention; determining multiple directed propagation paths of the intervention variable based on the connection relationships between nodes in the pre-defined causal network; the directed propagation path starts at the node corresponding to the intervention variable and ends at the node corresponding to the target outcome; for each directed propagation path, according to the... The risk value of the target outcome under the intervention state is determined by the conditional probability, node state probability, and contribution to the risk of the target outcome of each node in the directed propagation path. Counterfactual state risk information of the directed propagation path is generated, and counterfactual effect information of the directed propagation path is generated by combining the factual state risk information. Each directed propagation path is ranked according to the counterfactual effect information of the intervention variable and the counterfactual effect information of the directed propagation path, and at least one critical path is identified according to the ranking results. The risk assessment result of the intervention variable is generated based on the factual state risk information, the counterfactual state risk information and counterfactual effect information of the intervention variable, and at least one critical path.

[0011] Optionally, the perinatal birth defect risk assessment method further includes: based on the risk assessment results of each intervention variable, traversing each directed propagation path in the preset causal network to determine the target path tree; the target path tree is used to characterize the effective intervention process for the target object; in the preset causal network, the target path tree is highlighted to generate a risk network diagram, and the risk network diagram is displayed to the target object.

[0012] According to a second aspect of the present invention, a perinatal birth defect risk assessment device is provided, comprising: an information collection module, configured to, in response to individual multidimensional data input by a target object, map the individual multidimensional data into individual variable information corresponding to a preset causal network, and determine observed variables and missing variables; a missing variable analysis module, configured to, based on the observed variables and the preset causal network, perform posterior inference on the missing variables to obtain a posterior distribution corresponding to the missing variables; a factual risk module, configured to, based on the observed variables and the posterior distribution, infer the risk value of the target outcome and generate factual state risk information; a counterfactual risk module, configured to, after determining the intervention variable according to the individual variable information and a preset set of interventionable variables, perform counterfactual intervention on the intervention variable, and, based on the observed variables and the posterior distribution, infer the risk value of the target outcome under the intervention state and generate counterfactual state risk information of the intervention variable; and a result generation module, configured to, based on the preset causal network, and according to the factual state risk information and the counterfactual state risk information of the intervention variable, generate a risk assessment result of the intervention variable.

[0013] According to a third aspect of the present invention, an electronic device is provided, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of the first aspect or any optional embodiment of the first aspect.

[0014] By adopting the above technical solution, this invention introduces counterfactual intervention to achieve quantitative prediction of the intervention effect, and at the same time, it uses causal networks to handle complex dependencies and data missing problems, making the evaluation results more reliable. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the perinatal birth defect risk assessment method in an embodiment of the present invention. Figure 2 This is a schematic diagram of the risk network diagram in an embodiment of the present invention; Figure 3 This is a schematic diagram of the perinatal birth defect risk assessment device in an embodiment of the present invention; Figure 4 This is a specific example diagram of an electronic device in an embodiment of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0018] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0019] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0020] Before providing a further detailed description of the embodiments of the present invention, the nouns and terms involved in the embodiments of the present invention will be explained, and the nouns and terms involved in the embodiments of the present invention shall be interpreted as follows.

[0021] A pre-constructed causal network refers to a probabilistic graphical model that represents the probabilistic dependencies between multiple risk factor variables and the target outcome of birth defects. In a preferred embodiment of the present invention, this network can be a directed acyclic graph (DAG), such as a Bayesian network. The network uses nodes to represent variables (such as maternal age, pre-pregnancy folic acid supplementation status, etc.), directed edges to represent causal or correlational relationships between variables, and stores conditional probability parameters quantifying these relationships.

[0022] Individual multidimensional data refers to a dataset related to a specific target individual and encompassing information across multiple dimensions. For example, it may include the individual's sociodemographic characteristics, lifestyle habits, environmental exposures, past obstetric history, history of chronic diseases, physical examination results, laboratory test results, and ultrasound examination results at different stages of pregnancy (pre-pregnancy, early pregnancy, mid-pregnancy, and / or late pregnancy). This data forms the basis for individualized risk assessment.

[0023] Posterior inference: In probabilistic graphical models, this involves calculating the conditional probability distribution of unobserved variables (missing variables) based on observed evidence (observed variables). This invention utilizes this process to handle incomplete target object data prevalent in the real world. By using known information to make probabilistic inferences about unknown information, it can generate robust risk assessment results even with incomplete data.

[0024] Counterfactual intervention: refers to a simulation process based on causal inference, that is, while keeping other conditions constant, a hypothetical state is forced on one or more interventionist variables in the model, and the possible changes in the target outcome under that hypothesis are inferred.

[0025] Critical Paths: In a pre-defined causal network, these are one or more directed propagation paths from an modifiable variable to the node corresponding to the target outcome. These paths contribute the most to the final risk probability and reveal the main transmission mechanisms of risk formation. Identifying these paths helps target subjects and physicians intuitively understand the sources of risk and find the most effective intervention targets.

[0026] To at least partially address one or more of the aforementioned problems and other potential issues, this invention proposes a perinatal birth defect risk assessment method. This method overcomes the limitations of existing risk assessment models, which can only output risk probabilities and cannot simulate the effects of interventions or provide quantitative decision-making guidance for target groups. By constructing and utilizing a pre-defined causal network, this method not only handles complex dependencies among multiple factors and data gaps, but more importantly, it introduces a counterfactual inference mechanism to achieve quantitative prediction of intervention effects, providing forward-looking and actionable decision support for target groups.

[0027] This invention provides a method for assessing the risk of birth defects during the perinatal period. Figure 1This is a flowchart illustrating a method for assessing the risk of birth defects during the perinatal period. This method can be applied to a perinatal birth defect risk assessment device. This device is located in an electronic device. The electronic device includes, but is not limited to, fixed and / or mobile devices. For example, fixed devices include, but are not limited to, servers, which can be cloud servers or regular servers. In some possible implementations, this perinatal birth defect risk assessment method can also be implemented by a processor calling computer-readable instructions stored in memory. Figure 1 As shown, this perinatal birth defect risk assessment method includes: S101. In response to the individual multidimensional data input by the target object, map the individual multidimensional data to individual variable information corresponding to the preset causal network, and determine the observed variables and missing variables.

[0028] S102. Based on the observed variables and the pre-defined causal network, perform posterior inference on the missing variables to obtain the posterior distribution corresponding to the missing variables.

[0029] S103. Based on the observed variables and posterior distribution, infer the risk value of the target outcome and generate factual state risk information.

[0030] S104. After determining the intervention variables based on individual variable information and the preset set of interventionable variables, counterfactual intervention is performed on the intervention variables. Based on the observed variables and the posterior distribution, the risk value of the target outcome under the intervention state is inferred, and the counterfactual state risk information of the intervention variables is generated.

[0031] S105. Based on the pre-set causal network, generate the risk assessment results of the intervention variables according to the factual state risk information and the counterfactual state risk information of the intervention variables.

[0032] In this embodiment of the invention, the target object can be a woman preparing for pregnancy, a pregnant woman, or a family member related to pregnancy outcome. The target object inputs individual multidimensional data via a mobile terminal, web page, or medical institution information system. This individual multidimensional data includes, but is not limited to, age, height, weight, past obstetric history, family history of genetic diseases, history of chronic diseases, medication history, folic acid supplementation status, smoking and drinking habits, infection status, environmental exposure status, prenatal examination indicators, and ultrasound screening results. Further, after receiving the individual multidimensional data input by the target object, the data can first be mapped to individual variable information corresponding to each node in a preset causal network according to a preset variable dictionary. For example, "not regularly supplementing folic acid before pregnancy" can be mapped to a value of the "folic acid supplementation status before pregnancy" node, and "history of viral infection in early pregnancy" can be mapped to a value of the "infection exposure in early pregnancy" node. Subsequently, based on the value requirements of each node in the preset causal network, nodes with valid values ​​can be identified as observed variables, while nodes that are not filled in, incompletely filled in, have abnormal values, or whose status cannot be directly determined can be identified as missing variables.

[0033] In this embodiment of the invention, for missing variables, posterior inference can be performed based on observed variables and a pre-defined causal network. For example, when the target subject has not filled in the "pre-pregnancy environmental pollutant exposure" information, the different states of the "pre-pregnancy environmental pollutant exposure" node can be probabilistically inferred by combining observed variables such as their occupation type, residential area, smoking history, and relevant examination indicators, thereby obtaining the posterior distribution of the missing variable. This posterior distribution can be represented as the probability of the missing variable being in each candidate state.

[0034] In this embodiment of the invention, after obtaining the posterior distributions of the observed and missing variables, the risk value of the target outcome can be inferred based on a preset causal network, thereby generating factual state risk information. The target outcome can be the overall risk of birth defects, or the risk of specific birth defect types such as neural tube defects, congenital heart disease, chromosomal abnormalities, and cleft lip and palate. The posterior distributions of the observed and missing variables in the current state of the target object can be input into the preset causal network to calculate the expected probability of occurrence corresponding to the node lock of the target outcome, and can be further converted into risk levels such as low risk, medium risk, or high risk.

[0035] In this embodiment of the invention, intervention variables can be determined based on individual variable information and a preset set of interventionable variables. For example, interventionable variables may include folic acid supplementation status, weight control status, blood glucose control status, smoking and alcohol exposure, infection treatment status, medication adjustment status, nutritional status, or environmental exposure status. Subsequently, counterfactual interventions can be performed on the determined intervention variables. For example, "no folic acid supplementation before pregnancy" can be set to "regular folic acid supplementation before pregnancy," or "poor blood glucose control in late pregnancy" can be set to "good blood glucose control in late pregnancy." The influence of the original parent node of the intervention variable on its status can be severed in a preset causal network, placing the intervention variable in the set intervention state. Finally, the posterior distributions of observed and missing variables can be combined to infer the risk value of the target outcome under the intervention state, thereby generating counterfactual risk information for the intervention variable.

[0036] In this embodiment of the invention, based on a pre-defined causal network, factual risk information can be compared with counterfactual risk information of intervention variables to generate risk assessment results for the intervention variables. Specifically, the risk assessment results for each intervention variable can be integrated to generate an overall risk assessment result. This overall risk assessment result may include the expected probability of the target outcome, risk level, main risk factors, modifiable factors, the magnitude of risk change before and after intervention, and personalized risk warning information.

[0037] The technical solution in this invention can still use a pre-set causal network to probabilistically complete missing variables and perform posterior inference even when the individual data of the target object is missing, incomplete, or uncertain. This avoids risk assessment bias caused by simply deleting missing samples or using fixed default values. Furthermore, it not only provides the factual risk of the target object's current state but also assesses the potential risk changes brought about by different intervention measures through counterfactual intervention reasoning. This gives the risk assessment results stronger individualized interpretability and intervention guidance value, facilitating early identification, early warning, and early intervention of birth defect risks during the perinatal period.

[0038] In some embodiments, the pre-constructed causal network is pre-built as follows: a set of candidate risk variables is determined based on documentary evidence, historical sample data, and expert knowledge; a constrained structural learning algorithm is used to perform conditional independence tests on the set of candidate risk variables to generate a network skeleton; temporal constraints are established according to the order of occurrence of variables at different stages of the perinatal period; edge constraints are established based on temporal constraints and expert knowledge; edge constraints include prohibition of edge constraints and / or mandatory edge constraints; the network structure is optimized using a scoring search algorithm based on the network skeleton, temporal constraints, and edge constraints to generate a target directed acyclic graph; training samples are constructed based on historical sample data, and parameter learning is performed on the target directed acyclic graph based on the training samples to obtain the conditional probability parameters corresponding to each node; the pre-constructed causal network is constructed based on the target directed acyclic graph and the conditional probability parameters.

[0039] In this embodiment of the invention, the pre-built causal network can be constructed before performing individual risk assessment. For example, candidate risk variables related to perinatal birth defects can be extracted from medical guidelines, epidemiological research literature, birth defect surveillance data, historical hospital sample data, and expert knowledge in fields such as reproductive medicine, obstetrics and gynecology, pediatrics, and public health, forming a set of candidate risk variables. Candidate risk variables may include maternal age, family genetic background, history of abnormal pregnancies, folic acid supplementation, nutritional status, body mass index, blood glucose level, blood pressure level, infection exposure, drug exposure, radiation exposure, environmental pollutant exposure, prenatal screening indicators, and abnormal ultrasound findings, etc.

[0040] In this embodiment of the invention, a constrained structural learning algorithm can be used to perform conditional independence tests on the set of candidate risk variables to determine whether there are statistical dependencies between the variables, and thereby generate an initial network skeleton. The network skeleton represents the undirected connections that may exist between the candidate risk variables.

[0041] In this embodiment of the invention, after generating the network skeleton, temporal constraints can be established based on the chronological order of variables at different stages of the peripregnancy period. For example, preconception factors precede early pregnancy factors, early pregnancy factors precede mid-pregnancy factors, and mid-pregnancy factors precede late pregnancy factors. This temporal constraint ensures that the causal direction does not violate medical time order. Furthermore, edge constraints can be established by incorporating expert knowledge. These constraints include prohibited and mandatory edge constraints. For example, an edge pointing from "maternal gestational diabetes" to "fetal chromosomal abnormality" can be set as a prohibited edge; an edge pointing from "maternal gestational diabetes" to "fetal developmental abnormality" can be set as a mandatory edge.

[0042] In this embodiment of the invention, based on the network skeleton, temporal constraints, and edge constraints, a scoring search algorithm can be used to optimize the network structure. For example, the scoring search algorithm can employ greedy equivalence search, hill-climbing search, etc., and can select the network structure with the better score as the target directed acyclic graph while satisfying the directed acyclic graph constraints.

[0043] In this embodiment of the invention, training samples can be constructed based on historical sample data to learn the parameters of the target directed acyclic graph. For example, for discrete variables, maximum likelihood estimation or Bayesian estimation can be used to obtain the conditional probability parameters of each node under different combinations of values ​​of its parent node; for historical sample data with missing values, the expectation-maximization algorithm can be used for parameter estimation. For continuous variables, medical threshold discretization can be performed first, or a conditional Gaussian model can be used for parameter learning. Finally, the target directed acyclic graph and the conditional probability parameters corresponding to each node can be combined to construct a preset causal network.

[0044] Thus, the construction of the pre-defined causal network integrates documentary evidence, historical sample data, and expert knowledge. It can utilize real sample data to discover statistical dependencies between variables, while avoiding the generation of network structures that violate medical common sense or have incorrect causal directions through temporal constraints and expert edge constraints. The resulting pre-defined causal network not only has good data fitting ability but also strong medical rationality and interpretability, providing a reliable model foundation for subsequent missing variable inference, factual risk assessment, and counterfactual intervention analysis.

[0045] In some embodiments, in response to the individual multidimensional data input by the target object, the individual multidimensional data is mapped to individual variable information corresponding to a preset causal network to determine observed variables and missing variables, including: in response to the individual multidimensional data input by the target object, the individual multidimensional data is validated to obtain processed individual data; the individual multidimensional data includes individual data corresponding to at least one stage of the target object in preconception, early pregnancy, mid-pregnancy and / or late pregnancy; based on a preset variable dictionary, each processed individual data is mapped to variable information of the corresponding node in the preset causal network to obtain a set of variable information that corresponds one-to-one with the individual multidimensional data; based on predefined conditions in the preset causal network, variables in the variable information set that meet the predefined conditions are determined as observed variables, and variables in the variable information set that do not meet the predefined conditions are determined as missing variables; the predefined conditions include at least one of integrity verification conditions, legal value range conditions, logical consistency conditions, and invalid identifier filtering conditions.

[0046] In this embodiment of the invention, after responding to the individual multidimensional data input by the target object, the individual multidimensional data can first be validated and preprocessed. The individual multidimensional data can come from questionnaires actively filled out by the target object, hospital electronic medical records, laboratory testing systems, prenatal screening systems, or data collected by wearable devices, covering one or more stages of preconception, early pregnancy, mid-pregnancy, and late pregnancy.

[0047] In this embodiment of the invention, integrity checks, format checks, value range checks, and logical consistency checks can be performed on individual multidimensional data. For example, it can determine whether age, gestational age, body mass index, blood glucose level, blood pressure level, etc., are within the valid value range; determine whether there is a logical conflict between "not pregnant" and "gestational age is 12 weeks"; determine whether "no history of medication use" is inconsistent with the specific name of the teratogenic risk drug filled in; and determine whether certain fields are filled in with invalid markers such as "unknown," "unknown," or "refused to answer." For data with format errors but correctable, unit conversion or standardization can be performed; for data whose validity status cannot be determined, it can be reserved as a pending inference state.

[0048] In this embodiment of the invention, after verification, the processed individual data can be mapped to node variable information in a preset causal network based on a preset variable dictionary. The variable dictionary may include original field names, standard variable names, variable types, value codes, medical thresholds, unit conversion rules, and synonym mapping rules. For example, "number of cigarettes smoked per day" and "secondhand smoke exposure time" can be mapped to the "tobacco exposure" node; "fasting blood glucose" and "glycated hemoglobin" can be mapped to the "glucose metabolism status" node; and "nuchal translucency scan value" and "structural ultrasound abnormality indication" can be mapped to the "prenatal screening abnormality" node. Each processed individual data point can be mapped to the variable information of the corresponding node, thereby forming a set of variable information corresponding to the individual's multidimensional data.

[0049] In this embodiment of the invention, variables in the variable information set can be classified according to predefined conditions in a preset causal network. If a variable meets the integrity verification condition, the legal value range condition, the logical consistency condition, and the invalid identifier filtering condition, it is determined as an observed variable; if a variable does not meet any of the above conditions, or if the variable lacks necessary values, it is determined as a missing variable. Thus, a set of observed variables and a set of missing variables suitable for subsequent causal inference are obtained.

[0050] In this way, individual data from different sources and in different formats during the perinatal period can be uniformly mapped to standard variable information that can be identified by a pre-defined causal network, avoiding the impact of problems such as inconsistent field naming, inconsistent units, and different value encoding on model inference. At the same time, through multiple checks such as completeness, legal values, logical consistency, and invalid identifier filtering, it is possible to effectively distinguish between real observation information and unreliable or missing information, improve the quality of input data, and thus improve the accuracy and stability of birth defect risk assessment results.

[0051] In some embodiments, based on observed variables and a preset causal network, posterior inference is performed on missing variables to obtain a posterior distribution corresponding to the missing variables, including: determining the candidate state range of the missing variables according to the relevant nodes of the observed variables in the preset causal network; the relevant nodes include at least neighboring nodes, ancestor nodes and temporally preceding nodes; performing posterior inference on the missing variables to obtain a posterior distribution; the posterior distribution includes a univariate posterior distribution and / or a joint posterior distribution.

[0052] In this embodiment of the invention, once a missing variable is identified, the candidate state range of the missing variable can be determined based on the relevant nodes of the observed variables in a preset causal network. For each missing variable, its neighboring nodes, ancestor nodes, and temporally preceding nodes in the preset causal network can be searched. Neighboring nodes can include the parent and child nodes of the missing variable, ancestor nodes can include upstream nodes that point to the missing variable through one or more directed edges, and temporally preceding nodes can include nodes that occur earlier than the stage at which the missing variable occurred during the perinatal period.

[0053] For example, when the variable "early pregnancy infection exposure" is missing, its candidate status range can be determined as "no infection exposure," "suspected infection exposure," and "confirmed infection exposure" by combining its adjacent nodes "early pregnancy fever symptoms," "early pregnancy anti-infective medication," the ancestral node "pre-pregnancy immune status," and the chronological preceding node "past infection history." As another example, when the variable "pre-pregnancy folic acid supplementation status" is missing, its candidate status range can be determined as "regular supplementation," "irregular supplementation," and "no supplementation" by combining information from the preconception stage, nutritional supplementation records, and prenatal checkup records.

[0054] In this embodiment of the invention, after determining the range of candidate states, posterior inference can be performed on the missing variables based on a preset causal network. Specifically, observed variables can be input as evidence variables into the preset causal network, and the posterior probability of the missing variable in each candidate state can be calculated based on the conditional probability parameters of each node in the preset causal network. Posterior inference can be implemented using variable elimination algorithms, cluster tree propagation algorithms, belief propagation algorithms, importance sampling algorithms, or Markov chain Monte Carlo sampling algorithms. For a single missing variable, a univariate posterior distribution can be output; for multiple interrelated missing variables, a joint posterior distribution can be further output to characterize the probability of combinations of states of multiple missing variables.

[0055] Thus, when inferring missing variables, the individualized information contained in neighboring nodes, ancestor nodes, and temporally preceding nodes related to the missing variables is fully utilized, making the inference results of missing variables more consistent with the actual situation of the target object. At the same time, univariate posterior distributions and joint posterior distributions can preserve the uncertainty of missing information, providing probabilistic inputs for subsequent factual risk calculation and counterfactual risk calculation, and avoiding underestimation of uncertainty in risk assessment due to simply filling in a single value.

[0056] In some embodiments, the risk value of the target outcome is inferred based on the observed variables and the posterior distribution, and factual state risk information is generated, including: calculating the risk value of the target outcome in each candidate state by weighting multiple candidate states in the posterior distribution according to the observed variables and the posterior probability; summarizing the risk values ​​in each candidate state by weight to obtain the risk value of the target outcome, and generating factual state risk information.

[0057] In this embodiment of the invention, after obtaining the posterior distributions of the observed and missing variables, factual state risk reasoning can be performed on the target outcome. For example, the observed variables can first be input as fixed evidence into a preset causal network, and the risk values ​​of the target outcome under different candidate states can be weighted and calculated based on the posterior probabilities of multiple candidate states in the posterior distribution of the missing variables.

[0058] For example, if the posterior distribution of the missing variable "pre-pregnancy folic acid supplementation status" is 0.30 for "regular supplementation", 0.50 for "irregular supplementation", and 0.20 for "no supplementation", then the risk values ​​can be calculated for each of the three candidate states. If the target outcome is the risk of neural tube defects, the expected probability of neural tube defects can be calculated for each of the "regular supplementation", "irregular supplementation", and "no supplementation" states.

[0059] Specifically, for any candidate state in the posterior distribution, this candidate state can be treated as a possible value of the missing variable and input into a pre-defined causal network along with the observed variables as evidence. The expected probability of the target outcome occurring in this candidate state is then inferred and used as the risk value corresponding to that candidate state. In particular, the risk value can be calculated separately for each candidate state. The posterior probability corresponding to each candidate state is used to characterize the likelihood of that candidate state occurring given the current individual information of the target object, and can be used as a weighting factor in subsequent risk aggregation.

[0060] Furthermore, for cases with multiple missing variables, multiple candidate state combinations can be determined based on the joint posterior distribution of the missing variables. For any candidate state combination, the values ​​of each missing variable in the candidate state combination, along with the observed variables, can be input into a pre-defined causal network to calculate the risk value of the target outcome under that candidate state combination; simultaneously, the joint posterior probability corresponding to the candidate state combination is used as the weighted weight of the risk value. If the multiple missing variables are independent or approximately independent, the combination probability of the candidate state combinations can also be determined based on the product of the posterior probabilities of each missing variable.

[0061] In this embodiment of the invention, after calculating the risk value for each candidate state or combination of candidate states, the risk values ​​can be weighted and aggregated according to the corresponding posterior probabilities to obtain the risk value of the target outcome, thereby generating factual state risk information. This factual state risk information may include the expected probability of occurrence of the target outcome, its confidence interval, risk level, and relative risk level compared to the baseline risk of the population. For example, the weighted expected probability of occurrence can be compared with a preset risk threshold; if the expected probability of occurrence is higher than the risk threshold, a high-risk warning is output; if it is comparable to the risk threshold, a medium-risk warning is output.

[0062] Thus, the process of generating fact-state risk information can fully consider the impact of missing variables in different states on the risk of the target outcome, and perform weighted summation based on the posterior probabilities of each state. The resulting fact-state risk is not based on a single assumed state, but comprehensively reflects the overall risk level of the target subject under the current information conditions, which can improve the robustness and credibility of risk assessment results, and help medical staff or target subjects understand the current source of risk.

[0063] In some embodiments, after determining the intervention variable based on individual variable information and a preset set of interventionable variables, counterfactual intervention is performed on the intervention variable, and the risk value of the target outcome under the intervention state is inferred based on the observed variables and the posterior distribution, generating counterfactual state risk information of the intervention variable. This includes: determining at least one intervention variable from the preset set of interventionable variables based on individual variable information; setting the intervention variable as the target intervention state according to preset intervention rules, and suppressing the influence of the original causal input of the intervention variable on its state in a preset causal network; calculating the risk value of the target outcome under each candidate state by weighting multiple candidate states in the posterior distribution according to the observed variables and the intervention variable and the posterior probability; and summarizing the risk values ​​under each candidate state by weight to obtain the risk value of the target outcome, thereby generating counterfactual state risk information of the intervention variable.

[0064] In this embodiment of the invention, a set of modifiable variables can be pre-set. The set of modifiable variables includes variables that can be changed in perinatal management through medical guidance, lifestyle adjustment or environmental control, such as folic acid supplementation status, nutritional intake status, blood glucose control status, blood pressure control status, weight management status, smoking and drinking exposure, drug use status, infection treatment status, occupational and environmental exposure protection status, etc.

[0065] In this embodiment of the invention, at least one intervention variable can be determined from a preset set of interventionable variables based on the individual variable information of the target object. Exemplarily, the process of determining the intervention variable can be random selection or selection according to preset rules; this invention does not limit this. In particular, since missing variables can also be assigned a target intervention state and counterfactually estimated, the intervention variable can be either an observed variable or a missing variable.

[0066] In this embodiment of the invention, after determining the intervention variable, it can be set as the target intervention state according to a preset intervention rule. For example, the preset intervention rule can be randomly selected or determined based on medical guidelines or expert knowledge. For instance, "irregular folic acid supplementation before pregnancy" can be set as "regular folic acid supplementation before pregnancy" or "no folic acid supplementation before pregnancy," thereby calculating the individualized effect after the change in intervention state. Furthermore, in the preset causal network, a counterfactual intervention can be performed on the intervention variable, that is, fixing the intervention variable as the target intervention state and suppressing the influence of its original causal inputs on its state, so that upstream causes no longer change the value of the intervention variable.

[0067] In this embodiment of the invention, inference can be performed based on the posterior distributions of observed variables, intervention variables, and missing variables under counterfactual intervention conditions. For each candidate state or combination of candidate states in the posterior distribution of the missing variable, the risk value of the target outcome under the intervention state is calculated, and then weighted and summarized according to the posterior probabilities corresponding to the candidate states or combinations of candidate states to obtain the risk value of the target outcome, thereby generating counterfactual state risk information for the intervention variable. This counterfactual state risk information may include the expected probability of occurrence of the target outcome under the intervention state, the confidence interval, the risk level, and the relative risk level compared to the baseline risk of the population.

[0068] In this way, it is possible to assess the risk level of the target group based on their individual characteristics and quantify the potential risk changes brought about by different intervention measures through counterfactual intervention reasoning. At the same time, when calculating the counterfactual state risk information, the posterior uncertainty of the missing variables is still retained, and the risk values ​​of different candidate states are weighted and summarized according to the posterior probability, making the obtained counterfactual risk results more robust and consistent with real-world application scenarios. This is conducive to providing personalized and interpretable auxiliary decision-making suggestions for the perinatal risk management of the target group.

[0069] In some embodiments, based on a pre-defined causal network, and according to factual state risk information and counterfactual state risk information of the intervention variable, a risk assessment result for the intervention variable is generated, including: generating counterfactual effect information of the intervention variable based on factual state risk information and counterfactual state risk information of the intervention variable; the counterfactual effect information of the intervention variable includes the risk difference, risk ratio, and / or risk change magnitude of the target outcome before and after the intervention; determining multiple directed propagation paths of the intervention variable based on the connection relationships between nodes in the pre-defined causal network; the directed propagation path starts at the node corresponding to the intervention variable and ends at the node corresponding to the target outcome; for each directed propagation path, based on... Based on the conditional probability, node state probability, and contribution to the risk of the target outcome of each node in the directed propagation path, the risk value of the target outcome under the intervention state is determined, and counterfactual state risk information of the directed propagation path is generated. Combined with the factual state risk information, counterfactual effect information of the directed propagation path is generated. Each directed propagation path is ranked according to the counterfactual effect information of the intervention variable and the counterfactual effect information of the directed propagation path, and at least one critical path is identified according to the ranking results. Based on the factual state risk information, the counterfactual state risk information and counterfactual effect information of the intervention variable, and at least one critical path, the risk assessment result of the intervention variable is generated.

[0070] In this embodiment of the invention, counterfactual effect information of the intervention variable can be generated based on factual risk information and counterfactual risk information of the intervention variable. This counterfactual effect information may include the risk difference, risk ratio, and / or the magnitude of risk change of the target outcome before and after the intervention, and may also include other parameters or content that can reflect the risk change before and after the intervention; this invention does not limit this. Specifically, the magnitude of risk change can represent either the increase or decrease in risk, thereby prompting the target population and medical personnel to choose interventions with lower risk or higher intervention benefits.

[0071] In this embodiment of the invention, multiple directed propagation paths can be searched based on the connection relationships between nodes in a pre-defined causal network, starting from the intervention variable and ending at the node corresponding to the target outcome. A directed propagation path represents a possible chain of transmission where the intervention variable influences the target outcome step-by-step along the causal direction. For example, "assisted reproductive pregnancy—multiple pregnancy—birth defects" can constitute one directed propagation path; "lack of regular folic acid supplementation before pregnancy—insufficient folic acid levels in red blood cells during early pregnancy—neural tube defects" can constitute another directed propagation path.

[0072] In this embodiment of the invention, counterfactual state risk information for each directed propagation path can be calculated under intervention conditions, and combined with factual state risk information to generate counterfactual effect information for that directed propagation path. For example, counterfactual effect information can be calculated using the difference between counterfactual state risk information and factual state risk information. Further, counterfactual state risk information can be calculated based on the conditional probability, node state probability, and contribution to the target outcome risk of each node on the path. Here, conditional probability refers to the probability that a node in a pre-defined causal network is in a certain state when its parent node is in a specific state; this can be derived from a conditional probability table or conditional probability parameters obtained from training the pre-defined causal network, used to express the dependency relationship or risk transmission relationship between nodes. Node state probability refers to the probability that a node is in a certain state given the current individual information of the target object; this can come from the actual input of the target object, or can be obtained based on observed variables and inference from the pre-defined causal network, used to express the probability that a specific node is currently in a certain state. The contribution to the target outcome risk characterizes the degree of influence of the node in the formation of the target outcome risk, and can be determined through medical guidelines or expert knowledge. In particular, for paths containing missing variables, the path risk value can be calculated by combining the posterior distribution of the missing variables. The path risk value is used to characterize the expected probability of the target outcome under the risk transmission relationship defined by the directed propagation path.

[0073] In this embodiment of the invention, the direction of risk change before and after intervention can be determined based on the counterfactual effect information of the intervention variable. The direction of risk change can include risk increase, risk decrease, or no change in risk. Further, when the direction of risk change is risk increase, the counterfactual effect information of each directed propagation path can be arranged in descending order, and the path with positive counterfactual effect information can be selected as the critical path; conversely, when the direction of risk change is risk decrease, the counterfactual effect information of each directed propagation path can be arranged in ascending order, and the path with negative counterfactual effect information can be selected as the critical path.

[0074] In this embodiment of the invention, a risk assessment result for the intervention variable can be generated based on factual risk information, counterfactual risk information values ​​and counterfactual effect information of the intervention variable, and at least one critical path. Specifically, the risk assessment result may also include risk factors involved in the critical path, path propagation relationships, counterfactual risk information and counterfactual effect information corresponding to the path, and an explanation of the critical path.

[0075] Thus, the risk assessment results not only provide the overall risk of the target outcome, but also reveal the process by which the risk is transmitted from the source factors to the target outcome in the form of causal paths, making the source of the risk clearer and more interpretable. By sorting multiple directed transmission paths and identifying key paths, the most important risk chains of intervention factors can be highlighted, helping medical staff quickly grasp key risk factors and facilitating the target population to understand the causes of risk formation and the focus of intervention.

[0076] In some embodiments, the perinatal birth defect risk assessment method further includes: based on the risk assessment results of each intervention variable, traversing each directed propagation path in a preset causal network to determine a target path tree; the target path tree is used to characterize the effective intervention process for the target object; in the preset causal network, the target path tree is highlighted to generate a risk network diagram, and the risk network diagram is displayed to the target object.

[0077] In this embodiment of the invention, at least one intervention variable and its risk assessment result can be obtained first. Furthermore, since there may be nested, overlapping, or combined relationships between the directed propagation paths corresponding to different intervention variables, the directed propagation paths related to each intervention variable in the preset causal network can be traversed to identify shared nodes, upstream and downstream relationships, and convergence relationships between different paths.

[0078] In this embodiment of the invention, during the traversal process, the causal chain from each intervention variable to the target outcome can be traced backward from the node corresponding to the target outcome; alternatively, each intervention variable can be used as a starting point to search forward for all directed propagation paths that can reach the node corresponding to the target outcome. For multiple paths obtained through the search, the intervention effectiveness of each path is judged based on the corresponding counterfactual effect information. When the risk reduction of a path corresponding to a certain intervention variable reaches a preset condition, or when the combination of multiple intervention variables has a greater risk reduction effect than a single intervention variable, the corresponding path and node can be included in the target path tree. Simultaneously, when a certain intervention variable or a combination of multiple intervention variables has a risk reduction effect but a low risk reduction magnitude, the corresponding path and node can be used as a non-target path tree. Furthermore, when a certain intervention variable or a combination of multiple intervention variables does not have a risk reduction effect or the risk reduction magnitude is negligible, the corresponding path and node can be used as other path trees.

[0079] For example, the modifiable variables identified for the target population—"preconception obesity," "unhealthy diet during pregnancy," and "gestational diabetes mellitus"—are all associated with fetal developmental abnormalities. Specifically, "preconception obesity" and "unhealthy diet during pregnancy" indirectly affect fetal developmental abnormalities by influencing "gestational diabetes mellitus," while "gestational diabetes mellitus" directly affects fetal developmental abnormalities. Intervening only in "preconception obesity" (preconception weight loss) or only in "unhealthy diet during pregnancy" (healthy diet during pregnancy) fails to effectively control gestational blood sugar, a core teratogenic factor, resulting in a limited reduction in the risk of fetal developmental abnormalities. Intervening only in "gestational diabetes mellitus" (nutritional or pharmacological treatment) fails to block the cause early before or during pregnancy, achieving front-end prevention and only reducing some risks. However, simultaneously intervening in "preconception obesity," "unhealthy diet during pregnancy," and "gestational diabetes mellitus" yields the largest counterfactual risk reduction. The paths "preconception obesity—gestational diabetes mellitus—fetal developmental abnormalities," "unhealthy diet during pregnancy—gestational diabetes mellitus—fetal developmental abnormalities," and "gestational diabetes mellitus—fetal developmental abnormalities" can be combined to form a target path tree. This target path tree can represent the effective intervention process for the target group, namely, that controlling pre-pregnancy weight, pregnancy diet and gestational blood sugar can better weaken the transmission of fetal developmental abnormality risk.

[0080] Specifically, when determining the target path tree, different path combinations can be screened according to preset path combination screening rules. These rules may include: whether the overall risk reduction corresponding to the path combination is greater than the risk reduction of a single path; whether each intervention variable in the path combination is medically feasible; whether the path combination involves duplicate or ineffective interventions; whether the path combination covers the main risk transmission nodes of the target outcome; and whether the path combination contains key convergence nodes. Path combinations with significant counterfactual effects, a moderate number of intervention variables, high medical feasibility, and coverage of the main risk transmission chains can be prioritized as the target path tree.

[0081] In this embodiment of the invention, the target path, intervention variables, intermediate nodes, and nodes corresponding to the target outcome contained in the target path tree can be highlighted in a preset causal network to generate a risk network diagram. Highlighting methods may include thickening the edges in the target path tree, changing the color, adding borders or shadow effects, adding intervention labels to the intervention variable nodes in the target path tree, and annotating the risk difference, risk ratio, and / or risk reduction magnitude before and after intervention next to the corresponding path. For cases where there are multiple intervention variables in the target path tree, the system can also use different colors or numbers to distinguish different intervention starting points and display the combination relationship of multiple intervention paths through convergence nodes. Specifically, the target path tree can be displayed with a bold red line, while non-target path trees can be displayed in orange, and other path trees can be displayed in gray.

[0082] In this embodiment of the invention, after generating the risk network diagram, it can be displayed to target individuals or medical personnel via mobile terminals, web pages, doctor workstations, or health management platforms. Target individuals can view key nodes, target paths, and corresponding explanations in the risk network diagram, and can also click on nodes or paths to view the source, meaning, probability of risk, and suggested intervention measures for that risk factor.

[0083] This approach enables a comprehensive traversal and combined analysis of directed propagation paths corresponding to multiple intervention variables, identifying nesting, intersection, and convergence relationships between paths to form a target path tree representing the effective intervention process. Simultaneously, it more accurately reflects the risk transmission mechanism under the combined effects of multiple factors, avoiding the omission of key intervention variables with synergistic effects. Furthermore, by highlighting the target path tree, target subjects and healthcare professionals can intuitively understand the correlation between recommended interventions and their impact on birth defect risk, improving the interpretability of risk assessment results and the value of intervention guidance.

[0084] In some implementations, the core of this perinatal birth defect risk assessment method is a pre-constructed causal network for birth defects, a Bayesian network whose structure and parameters are learned from large-scale real-world data. Target participants input their personal information through a questionnaire interface on a smartphone, personal computer, or other terminal running a webpage or application. Figure 2 A schematic diagram of the risk network diagram in an embodiment of the present invention is shown, such as... Figure 2 As shown, the risk network graph contains multiple risk factor nodes and target outcome nodes, and the directed edges between nodes represent the probabilistic dependencies between nodes.

[0085] For example, the personal information entered by the target audience on the questionnaire page may include: a 34-year-old woman planning to conceive naturally, with no family history of birth defects, but whose spouse has a history of smoking. If the target audience does not fill in certain items (such as family income), those items are identified as missing variables.

[0086] Furthermore, based on the observed variables, posterior inference can be performed on the missing variables to calculate their probability distribution. Then, a probability propagation algorithm is executed to integrate the posterior distributions of the observed and missing variables throughout the network, calculating the posterior probability of the final "birth defect" node, which represents the current factual risk information of the target object.

[0087] Furthermore, if the target individual wishes to understand the impact of "if the spouse quits smoking" on the risk, they can issue a modification command through the terminal's interactive interface. Upon receiving this command, a counterfactual intervention process is initiated. Using causal inference algorithms such as marginal structure models, the state of the "father actively smoking" node is forcibly set to "no" in the preset causal network. Under this assumption, the risk of birth defects is recalculated, resulting in a hypothetical counterfactual risk information, denoted as risk probability p0. Finally, the current risk p and the hypothetical risk p0 are displayed to the target individual (e.g., the risk decreases from 2.5% to 2.1%). In particular, for modification commands actively issued by the target individual, the corresponding path can be highlighted, such as highlighting the critical path from "father actively smoking" to "birth defects." In this way, the target individual not only receives an individualized risk assessment but also, through interactive simulation, quantifies the health benefits of changing specific interventionable behaviors, thus obtaining clear and actionable guidance for preparing for pregnancy.

[0088] In some implementations, a pre-defined causal network can be constructed based on medical evidence and large-scale birth cohort data. For example, an umbrella review identifies potential risk factors for birth defects, and desensitized perinatal characteristics and pregnancy outcome data from 132,386 pregnant women are used as historical sample data to learn the structure and parameters of the pre-defined causal network. The umbrella review systematically searches Chinese and English medical databases, filters systematic reviews and / or meta-analyses related to birth defect risk factors, extracts information such as risk factors, outcomes, sample size, effect size, and confidence intervals, and estimates the pooled effect size of each risk factor on each birth defect outcome using a fixed-effects model or a random-effects model based on heterogeneity. This results in a set of candidate variables including sociodemographic characteristics, lifestyle habits and environmental exposures, preconception health and reproductive factors, and pregnancy health.

[0089] Subsequently, historical sample data can be substituted into the network, and a maximum-minimum parent-child node discovery algorithm can be used to construct a skeleton graph for different variable dimensions to identify candidate parent-child nodes with conditional dependencies on birth defects. Then, a hill-climbing search algorithm combined with Bayesian information criteria is used to score the network structure. By adding, deleting, or reversing edges, directed acyclic graphs for each dimension are obtained without creating directed cycles. These directed acyclic graphs can be further integrated into a multi-dimensional causal network for birth defects, and the direction of cross-dimensional connections can be determined based on the temporal relationships between variables and medical knowledge. For example, sociodemographic feature nodes can point to lifestyle and environmental exposure nodes, pre-pregnancy health status nodes, and pregnancy health status nodes; pre-pregnancy health status nodes can point to pregnancy health status nodes. At the same time, connections that do not conform to medical common sense, such as later-point-in-time factors influencing earlier-point-in-time factors or variable environmental factors influencing congenital genetic factors, can be avoided. Furthermore, experts in obstetrics and gynecology, pediatrics, and public health can be invited to demonstrate the rationality of the causal chain, the harmfulness of risk factors, and the operability, in order to revise and confirm the final pre-designed causal network structure.

[0090] In an exemplary network structure, maternal age, paternal education level, maternal history of diabetes, family history of birth defects, and fetal deafness gene screening results can serve as root nodes; family income, paternal active smoking, maternal work environment pollution, maternal reproductive history, maternal history of birth defects during pregnancy, maternal pregnancy method, maternal number of fetuses in this pregnancy, maternal folic acid intake before pregnancy, maternal threatened miscarriage in early pregnancy, fetal non-invasive prenatal genetic testing results, maternal gestational diabetes in mid-pregnancy, maternal threatened miscarriage in mid-pregnancy, and maternal gestational diabetes in late pregnancy can serve as intermediate nodes; birth defects are the nodes corresponding to the target outcome. Based on historical sample data, the network is trained to obtain conditional probability tables or transition probability parameters for each node.

[0091] Furthermore, the nodes corresponding to the target outcome can not only be set as overall birth defects, but also sequentially replaced with nodes representing birth defects in different systems, such as the nervous system, eyes, ears, nose, mouth, and nose, the circulatory system, the respiratory system, cleft lip and palate, the digestive system, the reproductive system, the urinary system, the musculoskeletal system, and chromosomal abnormalities, and parameter learning can be performed separately for each. Specifically, the nodes corresponding to the target outcome can also be set as severity grading nodes for mild, severe, or fatal birth defects to output risk information for birth defects of different severity levels.

[0092] After the target subject inputs individual data, observed variables can be input as evidence variables into the pre-defined causal network. The content of individual data collection for the target subject can be determined based on the target subject's current state; when the target subject is in a state of planning pregnancy, information such as planned pregnancy age, spouse's education level, family income, spouse's smoking status, work environment pollution status, history of diabetes, reproductive history, history of birth defects, family history of birth defects, planned pregnancy method, and pre-pregnancy folic acid intake plan are collected; when the target subject is pregnant, information such as actual pregnancy age, spouse's education level, family income, spouse's smoking status, work environment pollution status, history of diabetes, reproductive history, history of birth defects, family history of birth defects, actual pregnancy method, number of fetuses, pre-pregnancy folic acid intake, current gestational age, and pregnancy outcome are collected; if in the early stages of pregnancy (conception - 13 days)... +6 Based on the above information, further information will be collected regarding folic acid intake during early pregnancy and threatened miscarriage during early pregnancy; if in the second trimester (14-27 weeks)... +6 Based on the above information, further data collection will be conducted on non-invasive prenatal genetic testing results, deafness gene screening results, threatened miscarriage status in the second trimester, and gestational diabetes status in the second trimester. If the patient is in the third trimester (28 weeks to delivery), further data collection will be conducted on threatened preterm labor status and gestational diabetes status in the third trimester. For logically related questions, skip-answer control can be implemented based on prior answers. For example, questions about birth defects can be skipped if the target subject has no history of childbirth, and questions about multiple pregnancies assisted by reproductive technology can be skipped if the target subject plans natural pregnancy. The questionnaire results will be standardized into individual multidimensional data, and information that is not filled in, not tested, or cannot be confirmed will be added to the individual multidimensional data as missing variables.

[0093] Furthermore, individual multidimensional data can be mapped to node values ​​in a predefined causal network, and probabilistic propagation inference can be performed. Parent nodes pass causal support information to child nodes, and child nodes pass diagnostic support information to parent nodes. Each node updates its posterior probability by integrating upstream and downstream information. For missing variables, their possible states can be marginalized probabilistically, and different candidate states can be included in the target outcome risk calculation according to their posterior probabilities. When the target object modifies the state of any node on the webpage, the inference can be re-executed in real time to update the overall birth defect risk, the birth defect risk of each system, and the severity risk.

[0094] Furthermore, counterfactual risk simulations can be performed on modifiable variables. For example, for women under 35 years of age planning to conceive, the risk of birth defects at their current age is first calculated. If the target group chooses to simulate a scenario of "age 35 and above," the mother's age node is set to the corresponding state, and the counterfactual risk is re-inferred, thus highlighting the potential changes in risk associated with advanced maternal age. Similarly, for target groups who did not plan to regularly take folic acid before pregnancy or whose partners were exposed to smoking, the corresponding node can be adjusted to a state of regular folic acid use or no smoking exposure. The risk difference, risk ratio, or magnitude of risk change before and after the intervention can be calculated and displayed in conjunction with key risk transmission paths in a pre-defined causal network to assist the target group in perinatal self-management and medical decision-making.

[0095] This invention provides a perinatal birth defect risk assessment device, such as... Figure 3 As shown, the device may include: an information collection module 301, used to map individual multidimensional data input by the target object into individual variable information corresponding to a preset causal network, and determine observed variables and missing variables; a missing data analysis module 302, used to perform posterior inference on the missing variables based on the observed variables and the preset causal network, and obtain the posterior distribution corresponding to the missing variables; a fact risk module 303, used to infer the risk value of the target outcome based on the observed variables and the posterior distribution, and generate factual state risk information; a counterfactual risk module 304, used to determine the intervention variable based on the individual variable information and a preset set of interventionable variables, perform counterfactual intervention on the intervention variable, and infer the risk value of the target outcome under the intervention state based on the observed variables and the posterior distribution, and generate counterfactual state risk information of the intervention variable; and a result generation module 305, used to generate a risk assessment result of the intervention variable based on the preset causal network, the factual state risk information, and the counterfactual state risk information of the intervention variable.

[0096] In some embodiments, the pre-constructed causal network is pre-built as follows: a set of candidate risk variables is determined based on documentary evidence, historical sample data, and expert knowledge; a constrained structural learning algorithm is used to perform conditional independence tests on the set of candidate risk variables to generate a network skeleton; temporal constraints are established according to the order of occurrence of variables at different stages of the perinatal period; edge constraints are established based on temporal constraints and expert knowledge; edge constraints include prohibition of edge constraints and / or mandatory edge constraints; the network structure is optimized using a scoring search algorithm based on the network skeleton, temporal constraints, and edge constraints to generate a target directed acyclic graph; training samples are constructed based on historical sample data, and parameter learning is performed on the target directed acyclic graph based on the training samples to obtain the conditional probability parameters corresponding to each node; the pre-constructed causal network is constructed based on the target directed acyclic graph and the conditional probability parameters.

[0097] In some embodiments, the information collection module 301 is configured to: in response to the individual multidimensional data input by the target object, verify the individual multidimensional data to obtain processed individual data; the individual multidimensional data includes individual data corresponding to at least one stage of the target object in preconception, early pregnancy, mid-pregnancy and / or late pregnancy; based on a preset variable dictionary, map each processed individual data to variable information of a corresponding node in a preset causal network to obtain a variable information set that corresponds one-to-one with the individual multidimensional data; based on predefined conditions in the preset causal network, determine variables in the variable information set that meet the predefined conditions as observed variables, and determine variables in the variable information set that do not meet the predefined conditions as missing variables; the predefined conditions include at least one of integrity verification conditions, legal value range conditions, logical consistency conditions, and invalid identifier filtering conditions.

[0098] In some embodiments, the missing data analysis module 302 is configured to: determine the candidate state range of the missing variable based on the relevant nodes of the observed variable in a preset causal network; the relevant nodes include at least neighboring nodes, ancestor nodes and temporally preceding nodes; perform posterior inference on the missing variable to obtain the posterior distribution; the posterior distribution includes a univariate posterior distribution and / or a joint posterior distribution.

[0099] In some embodiments, the fact risk module 303 is used to: calculate the risk value of the target outcome in each candidate state by weighting multiple candidate states in the posterior distribution according to the observed variables and the posterior probability; and to calculate the risk value of the target outcome by weighting and summing the risk values ​​in each candidate state, thereby generating fact state risk information.

[0100] In some embodiments, the counterfactual risk module 304 is configured to: determine at least one intervention variable from a preset set of interventionable variables based on individual variable information; set the intervention variable as the target intervention state according to preset intervention rules, and suppress the influence of the original causal input of the intervention variable on its state in a preset causal network; calculate the risk value of the target outcome in each candidate state by weighting multiple candidate states in the posterior distribution according to the observed variables and the intervention variable and the posterior probability; and sum the risk values ​​in each candidate state by weighting to obtain the risk value of the target outcome, thereby generating counterfactual state risk information of the intervention variable.

[0101] In some embodiments, the result generation module 305 is configured to: generate counterfactual effect information of the intervention variable based on factual state risk information and counterfactual state risk information of the intervention variable; the counterfactual effect information of the intervention variable includes the risk difference, risk ratio, and / or risk change magnitude of the target outcome before and after intervention; determine multiple directed propagation paths of the intervention variable based on the connection relationships between nodes in a preset causal network; the directed propagation paths start from the node corresponding to the intervention variable and end at the node corresponding to the target outcome; for each directed propagation path, determine the risk value of the target outcome under the intervention state based on the conditional probability, node state probability, and contribution to the risk of the target outcome of each node in the directed propagation path, generate counterfactual state risk information of the directed propagation path, and generate counterfactual effect information of the directed propagation path in combination with the factual state risk information; sort each directed propagation path based on the counterfactual effect information of the intervention variable and the counterfactual effect information of the directed propagation path, and identify at least one critical path based on the sorting result; and generate a risk assessment result of the intervention variable based on the factual state risk information, the counterfactual state risk information of the intervention variable, the counterfactual effect information, and at least one critical path.

[0102] In some embodiments, the perinatal birth defect risk assessment device further includes: a path tree determination module 306 ( Figure 3 (Not shown in the image), used to traverse each directed propagation path in the pre-set causal network based on the risk assessment results of each intervention variable, to determine the target path tree; the target path tree is used to characterize the effective intervention process for the target object; risk map generation module 307 ( Figure 3 (Not shown in the image) is used to highlight the target path tree in a preset causal network, generate a risk network diagram, and display the risk network diagram to the target object.

[0103] The perinatal birth defect risk assessment device provided in this invention can still use a preset causal network to probabilistically complete missing variables and perform posterior inference even when the individual data of the target subject is missing, incomplete, or uncertain. This avoids risk assessment bias caused by simply deleting missing samples or using fixed default values. Furthermore, it not only provides the factual risk of the target subject in their current state but also assesses the potential risk changes brought about by different intervention measures through counterfactual intervention reasoning. This gives the risk assessment results stronger individualized interpretability and intervention guidance value, facilitating early identification, early warning, and early intervention of perinatal birth defect risks.

[0104] The acquisition, storage, and application of user personal information involved in the technical solution of this invention all comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0105] According to embodiments of the present invention, an electronic device is also provided. Figure 4 A schematic block diagram of an example electronic device 400 that can be used to implement embodiments of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0106] like Figure 4 As shown, device 400 includes a computing unit 401, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 402 or a computer program loaded from storage unit 408 into random access memory (RAM) 403. RAM 403 may also store various programs and data required for the operation of device 400. The computing unit 401, ROM 402, and RAM 403 are interconnected via bus 404. Input / output (I / O) interface 405 is also connected to bus 404.

[0107] Multiple components in device 400 are connected to I / O interface 405, including: input unit 406, such as keyboard, mouse, etc.; output unit 407, such as various types of monitors, speakers, etc.; storage unit 408, such as disk, optical disk, etc.; and communication unit 409, such as network card, modem, wireless transceiver, etc. Communication unit 409 allows device 400 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0108] The computing unit 401 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 401 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 401 performs the various methods and processes described above, such as the perinatal birth defect risk assessment method. For example, in some embodiments, the perinatal birth defect risk assessment method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 408. In some embodiments, part or all of the computer program can be loaded and / or installed on device 400 via ROM 402 and / or communication unit 409. When the computer program is loaded into RAM 403 and executed by the computing unit 401, one or more steps of the perinatal birth defect risk assessment method described above can be performed. Alternatively, in other embodiments, the computing unit 401 may be configured to perform a perinatal birth defect risk assessment method by any other suitable means (e.g., by means of firmware).

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for assessing the risk of birth defects during the perinatal period, characterized in that, include: In response to the individual multidimensional data input by the target object, the individual multidimensional data is mapped to individual variable information corresponding to a preset causal network to determine the observed variables and missing variables; Based on the observed variables and the preset causal network, posterior inference is performed on the missing variables to obtain the posterior distribution corresponding to the missing variables; Based on the observed variables and the posterior distribution, the risk value of the target outcome is inferred, and factual state risk information is generated. After determining the intervention variables based on the individual variable information and the preset set of interventionable variables, counterfactual intervention is performed on the intervention variables, and the risk value of the target outcome under the intervention state is inferred based on the observed variables and the posterior distribution, thereby generating counterfactual state risk information of the intervention variables; Based on the preset causal network, the risk assessment result of the intervention variable is generated according to the factual state risk information and the counterfactual state risk information of the intervention variable.

2. The method according to claim 1, characterized in that, The pre-constructed causal network is pre-built in the following manner: The set of candidate risk variables was determined based on literature evidence, historical sample data, and expert knowledge. A constrained structural learning algorithm is used to perform conditional independence tests on the candidate risk variable set to generate a network skeleton; Establish time constraints based on the order of occurrence of variables at different stages of the perinatal period; Based on the timing constraints and the expert knowledge, edge constraints are established; the edge constraints include prohibited edge constraints and / or mandatory edge constraints. Based on the network skeleton, the temporal constraints, and the edge constraints, a scoring search algorithm is used to optimize the network structure and generate the target directed acyclic graph. Training samples are constructed based on the historical sample data, and parameter learning is performed on the target directed acyclic graph based on the training samples to obtain the conditional probability parameters corresponding to each node. The preset causal network is constructed based on the target directed acyclic graph and the conditional probability parameters.

3. The method according to claim 1, characterized in that, The individual multidimensional data input in response to the target object is mapped to individual variable information corresponding to a preset causal network, and the observed variables and missing variables are determined, including: In response to the individual multidimensional data input by the target object, the individual multidimensional data is validated to obtain processed individual data; the individual multidimensional data includes individual data corresponding to at least one stage of the target object in preconception, early pregnancy, mid-pregnancy and / or late pregnancy. Based on a preset variable dictionary, each processed individual data is mapped to the variable information of the corresponding node in the preset causal network, resulting in a set of variable information that corresponds one-to-one with the individual multidimensional data. Based on the predefined conditions in the preset causal network, variables in the variable information set that meet the predefined conditions are identified as observed variables, and variables in the variable information set that do not meet the predefined conditions are identified as missing variables; the predefined conditions include at least one of integrity verification conditions, legal value range conditions, logical consistency conditions, and invalid identifier filtering conditions.

4. The method according to claim 1, characterized in that, The step of performing posterior inference on the missing variable based on the observed variable and the preset causal network to obtain the posterior distribution corresponding to the missing variable includes: Based on the relevant nodes of the observed variables in the preset causal network, the candidate state range of the missing variables is determined; the relevant nodes include at least neighboring nodes, ancestor nodes, and temporally preceding nodes; Perform posterior inference on the missing variable to obtain the posterior distribution; the posterior distribution includes a univariate posterior distribution and / or a joint posterior distribution.

5. The method according to claim 1, characterized in that, The process of inferring the risk value of the target outcome based on the observed variables and the posterior distribution, and generating factual state risk information, includes: Based on the observed variables, the risk value of the target outcome in each candidate state is determined by weighting multiple candidate states in the posterior distribution according to the posterior probability; The risk values ​​of each candidate state are weighted and summarized to obtain the risk value of the target outcome, thereby generating factual state risk information.

6. The method according to claim 1, characterized in that, After determining the intervention variable based on the individual variable information and a preset set of interventionable variables, counterfactual intervention is performed on the intervention variable. Based on the observed variables and the posterior distribution, the risk value of the target outcome under the intervention state is inferred, and counterfactual risk information of the intervention variable is generated, including: Based on the individual variable information, at least one intervention variable is determined from the preset set of interventionable variables; According to the preset intervention rules, the intervention variable is set as the target intervention state, and the influence of the original causal input of the intervention variable on its state is suppressed in the preset causal network; Based on the observed variables and the intervention variables, the risk value of the target outcome in each candidate state is determined by weighting multiple candidate states in the posterior distribution according to the posterior probability; The risk values ​​of each candidate state are weighted and summarized to obtain the risk value of the target outcome, thereby generating the counterfactual state risk information of the intervention variable.

7. The method according to claim 1, characterized in that, The process of generating a risk assessment result for the intervention variable based on the preset causal network, according to the factual state risk information and the counterfactual state risk information of the intervention variable, includes: Based on the factual state risk information and the counterfactual state risk information of the intervention variable, counterfactual effect information of the intervention variable is generated; the counterfactual effect information of the intervention variable includes the risk difference, risk ratio and / or risk change magnitude of the target outcome before and after the intervention; Based on the connection relationships between nodes in the preset causal network, multiple directed propagation paths for the intervention variable are determined; each directed propagation path starts at the node corresponding to the intervention variable and ends at the node corresponding to the target outcome. For each of the directed propagation paths, the risk value of the target outcome under the intervention state is determined based on the conditional probability, node state probability, and contribution to the risk of the target outcome of each node in the directed propagation path. Counterfactual state risk information of the directed propagation path is generated, and counterfactual effect information of the directed propagation path is generated by combining the factual state risk information. Each of the directed propagation paths is ranked according to the counterfactual effect information of the intervention variables and the counterfactual effect information of the directed propagation paths, and at least one critical path is identified based on the ranking results. Based on the factual state risk information, as well as the counterfactual state risk information, counterfactual effect information, and at least one critical path of the intervention variable, a risk assessment result for the intervention variable is generated.

8. The method according to claim 7, characterized in that, The method further includes: Based on the risk assessment results of each intervention variable, each directed propagation path in the preset causal network is traversed to determine the target path tree; the target path tree is used to characterize the effective intervention process for the target object. In the preset causal network, the target path tree is highlighted to generate a risk network graph, and the risk network graph is displayed to the target object.

9. A perinatal birth defect risk assessment device, characterized in that, include: The information collection module is used to respond to the individual multidimensional data input by the target object, map the individual multidimensional data into individual variable information corresponding to the preset causal network, and determine the observed variables and missing variables; The missing data analysis module is used to perform posterior inference on the missing variables based on the observed variables and the preset causal network, and obtain the posterior distribution corresponding to the missing variables; The fact risk module is used to infer the risk value of the target outcome based on the observed variables and the posterior distribution, and generate fact state risk information. The counterfactual risk module is used to determine the intervention variable based on the individual variable information and the preset set of interventionable variables, perform counterfactual intervention on the intervention variable, and infer the risk value of the target outcome under the intervention state based on the observed variables and the posterior distribution, and generate counterfactual state risk information of the intervention variable. The result generation module is used to generate a risk assessment result for the intervention variable based on the preset causal network, according to the factual state risk information and the counterfactual state risk information of the intervention variable.

10. An electronic device, characterized in that, include: At least one processor; as well as A memory that is communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by at least one processor to enable the at least one processor to perform the method of any one of claims 1-8.

Images