CN118335319A - Early prediction method for common major diseases based on virtual person simulation - Google Patents

Abstract

The invention belongs to the technical field of disease prediction, and discloses a method for early predicting common major diseases based on virtual human simulation, which comprises a virtual human prediction model based on a dynamic Bayesian network and constructed through structure learning and parameter learning, and early predicting the common major diseases through the virtual human prediction model; the invention combines multidimensional dynamic molecular change characteristics in the process of the longitudinal queue progress of the healthy age-increasing, builds a prediction model of 'virtual man' simulation based on the technologies of machine learning enhanced dynamic Bayesian network model, group learning and the like, reveals the 'multi-cause multi-fruit' combined effect of complex exposure and phenotype characteristics on various healthy outcomes, screens novel healthy age-increasing markers with biological significance, and improves the sensitivity and accuracy of early risk prediction of serious diseases in the process of the age-increasing through complementation among different layers of histology information.

Description

Early prediction method of common major diseases based on virtual human simulation

Technical Field

The present invention belongs to the technical field of disease prediction, and in particular relates to an early prediction method for common major diseases based on virtual human simulation.

Background technique

Common risk factors for major diseases include age, genetic factors, lifestyle, environmental factors, chronic disease history, psychological factors, etc. The factors affecting the onset of diseases cover a wide range and have relatively complex relationships. The analytical framework shown in Figure 2 provides us with a systematic method for evaluating the evidence of preventive medical services and provides important guidance for building an accurate risk assessment model. With the rapid development of multi-omics such as high-throughput genomics, proteomics, and metabolomics, it is possible to explain the complex mechanisms in the occurrence and development of common major diseases at the molecular level. Studies have shown that genetic factors also play an important role in the occurrence and development of common major diseases. Therefore, studying the risk of common major diseases at the molecular level and then building a multi-omics-based risk assessment model can help predict the phenotypes of common major diseases and stratify the risk of common major diseases, which will lay the foundation for precise prevention and diagnosis and treatment.

Traditional single-omics studies provide important information for screening new molecular markers of common major diseases in aging and the aging process, but the biological process information they provide often has great limitations, and the overall explanatory power of the complex aging process is also limited. The organic integration of information between different levels such as genes, proteins, metabolism, lipids, and microbiota not only provides more evidence to explain the biological mechanisms of the aging process, but also helps to deeply explore new molecular markers related to common major diseases in the aging process. Therefore, by integrating multi-dimensional information such as genomics, metabolomics, dietary patterns, and nutritional status, a risk assessment model for the onset of common major diseases is constructed to identify high-risk groups for common major diseases, dynamically assess the risk of individuals developing common major diseases in the future, and provide individualized early targeted preventive interventions for high-risk groups for common major diseases.

Summary of the invention

The purpose of the present invention is to provide an early prediction method for common major diseases based on virtual human simulation. The method adopts a dynamic Bayesian network model based on real causal relationships and machine learning as a virtual human prediction model, which can predict major diseases by stages and provide an important reference basis for risk assessment and prevention and treatment of common major diseases.

To achieve the above object, the present invention adopts the following technical solution:

A method for early prediction of common major diseases based on virtual human simulation, including a virtual human prediction model based on a dynamic Bayesian network constructed through structural learning and parameter learning, and early prediction of common major diseases through the virtual human prediction model;

The structural learning is used to screen out factors related to risk prediction from potential factors and determine the topological structure of the dynamic Bayesian network to construct a virtual human prediction model;

The structural learning is to first convert knowledge into normalized fuzzy membership by using fuzzy theory, determine the causal relationship between two variables by using the membership, and use the relationship between the two variables as a strong constraint or a weak constraint to search the Bayesian network structure space to construct multiple initial virtual human prediction models; then use the partitioned MCMC method to explore the optimal virtual human prediction model from the multiple initial virtual human prediction models;

The parameter learning is used to determine the probability distribution of each variable in the optimal virtual human prediction model under the condition of a given set of its parent nodes. The conditional probability relationship between the nodes in the optimal virtual human prediction model is estimated using the existing data set, and probability parameters are assigned to each node. The parameter learning uses the common major disease staging as the classification outcome indicator for risk assessment and first clusters through the EKM method and then uses the expanded two-stage stacking algorithm to predict the results. Finally, the optimal parameters of the virtual human prediction model are determined through the grid search method and 5-fold cross validation, so as to construct the final virtual human prediction model.

The amplified two-stage stacking algorithm uses the output class posterior probabilities of N different primary learners for the same data set as N fixed-dimensional input vectors of the meta-layer classifier, and adds a primary layer based on the stacking algorithm.

Furthermore, the causal relationship between two variables is determined by the membership degree: when the membership degree is 0, there is no causal relationship between the two variables; when the membership degree is 1, there is a causal relationship between the two variables; when the membership degree is between 0 and 1, the causal relationship between the two variables cannot be determined.

Furthermore, the strong restriction condition refers to the causal relationship between two variables when the membership degree is 0 and the membership degree is 1, and the weak restriction condition refers to the causal relationship between two variables when the membership degree is between 0 and 1.

Furthermore, the partitioned MCMC method comprises:

According to the partitioning requirements and partitioning rules, all variables in the initial topological structure of the Bayesian network are divided into m regions and numbered in sequence. Let the number of variables in the i-th region (i=1,2,…,m) be k _i , and the number of variables be , the number of variables in each partition , the specific variables in each partition are π _λ , the label partition Λ=(λ,π _λ ), and the Bayesian network structure under the label partition Λ is ; Given the data D, the posterior distribution of the label partition Λ is proportional to the total score obtained by merging the scores of each node _Xi and its parent node _Pai in the Bayesian network structure ,total score To verify the equivalence between the label partition space and the Bayesian network structure space, the optimal network structure in the Bayesian network structure space is determined based on the label partition with the largest total score;

In each iteration, the current marked partition is Λ, the proposed marked partition is Λ ^* , and the acceptance probability is ,in, To mark the partition The neighborhood of is partitioned by the label Split a partition into two partitions or merge two adjacent partitions into one partition.

Furthermore, the partitioning requirements include that there are no arrow connections between variables in the same zone, that variables in zone 1 have no parent nodes, and that each variable in each zone except zone 1 must have at least one parent node from the previous zone;

The partitioning rules are determined by strong constraints, and variables that are not clearly specified in the strong constraints are partitioned using a random simulation approach.

Furthermore, different primary learners are selected according to the type of data set. The primary learners for image data include convolutional neural network CNN, fully convolutional network FCN and deep Boltzmann machine DBM; the primary learners for text data include recurrent neural network RNN, long short-term memory network LSTM and gated recurrent unit GRU; the primary learners for numerical data include logistic regression, support vector machine SVM and naive Bayes.

The present invention combines the characteristics of multidimensional dynamic molecular changes in the process of healthy aging longitudinal cohort progress, and builds a prediction model for "virtual human" simulation based on machine learning-enhanced dynamic Bayesian network model and group learning and other technologies, revealing the "multi-cause and multi-effect" joint effects of complex exposure and phenotypic characteristics on multiple health outcomes, screening new markers of healthy aging with biological significance, and improving the sensitivity and accuracy of early risk prediction of major diseases in the aging process through the complementarity of omics information at different levels. On this basis, early risk screening tools suitable for clinical application are developed to promote the transformation and application of scientific research results and provide a scientific basis for the construction of a comprehensive prevention and treatment system for healthy aging.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic diagram of the process of the present invention.

Figure 2 shows the USPSTF analysis framework.

FIG3 is a schematic diagram of selecting a primary learner according to the present invention.

FIG. 4 is a comparison diagram of the model accuracy of the present invention.

FIG. 5 is a comparison diagram of model sensitivity of the present invention.

FIG6 is a comparison diagram of the AUC of the models of the present invention.

FIG. 7 is a comparison diagram of the model specificity of the present invention.

Detailed ways

As shown in Figure 1, this embodiment provides a method for early prediction of common major diseases based on virtual human simulation. The method builds a prediction model of virtual human simulation based on machine learning enhanced dynamic Bayesian network model and group learning technology, and effectively expresses and integrates multi-source information.

This embodiment constructs a dynamic Bayesian network as a virtual human prediction model through structure learning and parameter learning, and uses the virtual human prediction model to perform early prediction of common major diseases.

The structural learning is used to screen out factors related to risk prediction from a large number of potential factors, and determine the topological structure of the dynamic Bayesian network model to build a virtual human prediction model. This embodiment is dedicated to identifying factors that have an impact on the target variable, thereby building a more accurate and interpretable Bayesian network model structure; which involves analyzing and mining data to identify the correlation and causal relationship between variables; the results of structural learning will directly affect the quality and prediction performance of the virtual human prediction model.

This embodiment uses fuzzy theory and partitioned MCMC method to realize structural learning; there are two stages in structural learning. The first stage is knowledge-driven. Fuzzy theory is a mathematical tool for processing fuzzy and uncertain information, which can effectively deal with common fuzziness and uncertainty in the medical field; the goal of the first stage is to convert knowledge into normalized fuzzy membership. In short, it is hoped that the uncertain medical knowledge can be converted into a formalized representation through the fuzzification method for subsequent learning and reasoning processes; fuzzy membership can be regarded as a priori probability reflecting the causal relationship between variables; it helps to limit the number of variables and the complexity of the Bayesian network structure, thereby reducing the difficulty of structural learning to a certain extent. In the first stage, this embodiment uses rich knowledge and experience obtained from multiple sources such as medical databases and expert consultations; including but not limited to case data, medical literature, expert opinions, etc. Then, the rich knowledge and experience obtained are processed using fuzzy theory, and various information sources are comprehensively used to examine the causal relationship between these variables.

Generally speaking, the causal relationship between two variables A and B can be divided into three types: Type , A→B, that is, A is the cause of B; type , B→A, that is, B is the cause of A; type , there is no causal relationship between A and B.

These three types are used as three fuzzy subsets to measure the relationship between variables A and B, and the normalized fuzzy membership of the relationship between variables A and B belonging to the three fuzzy subsets is calculated based on each information source.

According to the size of the membership, it can be divided into the following three situations: , if the relationship between variables A and B is of a certain type, the fuzzy membership is equal to 0, indicating that variables A and B do not have such a relationship; for example, if the type of relationship between variables A and B is If the fuzzy membership is equal to 0, it means that A cannot be the cause of B. , if the relationship between variables A and B belongs to a certain type and the fuzzy membership is equal to 1, then variables A and B must have such a relationship; for example, if the type of relationship between variables A and B is If the membership degree is 1, then A must be the cause of B; If the relationship between variables A and B belongs to any type of membership between greater than 0 and less than 1, it means that it is difficult to determine the relationship between A and B through knowledge and experience, and further judgment is needed based on data characteristics.

For the situation and situation , which is used as a strong restriction condition for searching the Bayesian network structure space, that is, forcibly restricting the existence (or non-existence) of connection arrows between nodes A and B in the Bayesian network; for the above situation , it can be used as a weak restriction condition for searching the Bayesian network structure space and incorporated into the next stage of partitioned MCMC method structure learning in the form of prior probability; through strong and weak restriction conditions, the search range of the Bayesian network structure space can be narrowed, so as to simplify the task of searching the optimal Bayesian network structure by the partitioned MCMC method.

The second stage is to use the partitioned MCMC method to explore the optimal Bayesian network structure, using the Bayesian network as a model to describe the potential relationship between variables. The Bayesian network is a probabilistic graphical model that represents the dependencies between variables through a directed acyclic graph and uses probability distributions to describe these relationships. The goal of the second stage is to find the optimal network structure in the restricted Bayesian network structure space and use this network structure as the virtual human prediction model. The partitioned MCMC method can explore large search spaces more efficiently, thereby accelerating the learning process of the Bayesian network structure.

The partitioned MCMC method comprises the following steps:

First, all variables in the initial topological structure of the Bayesian network are divided into m regions and numbered in sequence, region 1, region 2, ... region m, and it is required that (a) there are no arrows connecting variables in the same region; (b) variables in region 1 have no parent nodes; (c) except region 1, each variable in each region must have at least one parent node from the previous region. The partitioning rules are mainly determined by the strong constraints in the first stage, and for variables that are not clearly specified in the strong constraints in the first stage, random simulation is used for partitioning.

Then, let the number of variables in the i-th region (i=1,2,…,m) be k _i , and the number of all variables , each partitioning method , using π _λ as the variable sorting corresponding to the partition method λ; according to the partitioning rule, λ is the number of variables in each partition, π _λ is to record the specific variables in each partition, mark the partition Λ=(λ,π _λ ), and use represents the Bayesian network structure under the label partition Λ; given the data D, the posterior distribution of the label partition Λ is proportional to the total score obtained by merging the scores of each node _Xi and its parent node _Pai in the Bayesian network structure ,total score The equivalence of the label partition space and the Bayesian network structure space is shown, that is, once the label partition with the largest score can be searched in the label partition space, the corresponding optimal network structure in the Bayesian network structure space can be determined based on the label partition with the largest score.

Then, we construct an MCMC method based on the label partition space. In each iteration of the MCMC method, we denote the current label partition as Λ, the proposed label partition as Λ ^* , and the acceptance probability as ,in, To mark the partition The neighborhood of is partitioned by the label This is achieved by splitting a partition into two partitions or merging two adjacent partitions into one partition. The partitioned MCMC method can provide a series of samples of Bayesian network structures that meet a specific stable posterior probability distribution under the condition of a given training data set, and further use Bayesian theory to perform statistical inference on these Bayesian network structure samples, so as to better deal with the uncertainty of the Bayesian network structure caused by high-dimensional complex data structures and sampling errors.

The structural learning of this embodiment is achieved through the combination of fuzzy theory and partitioned MCMC method, giving full play to the advantages of fuzzy theory and partitioned MCMC method in processing uncertain information; by limiting the search scope of the optimal Bayesian network structure to the scope of medical professional significance, it is ensured that the obtained structure has practical significance and can improve the efficiency of the algorithm.

The parameter learning is used to determine the probability distribution of each variable in the network given its parent node set. This embodiment will use existing data to estimate the conditional probability relationship between each node in the network, and assign appropriate probability parameters to each node in the Bayesian network, so that the network can more accurately reflect the statistical characteristics and probability distribution of the data. The basic construction process of the machine learning enhanced dynamic Bayesian network is constructed through structural learning and parameter learning, and the final virtual human prediction model is constructed; by combining real causal relationships and machine learning technology, the virtual human prediction model can better utilize multi-source information for risk prediction and decision support, and provide effective tools and methods for data analysis and prediction tasks applied in actual scenarios.

In the parameter learning process, this embodiment uses the staging of common major diseases as the classification outcome indicator for risk assessment; there are significant differences in the composition ratios of different categories of the classification outcome indicators of common major diseases. This embodiment combines the grading of different categories of common major diseases into one indicator, which further expands the differences in the composition ratios of different categories of composite indicators, making the data set unbalanced. When training disease prediction models in the face of unbalanced data sets, traditional machine learning algorithms usually tend to generate models that maximize the overall classification accuracy, while minority categories are easily ignored, which will lead to a decline in model performance, especially for minority groups. The prediction accuracy will be seriously affected. In view of the fact that the imbalance of data in different stages of common major diseases may affect the estimation of model parameters and reduce the running time of the algorithm, this embodiment follows the idea of decision-level fusion. After the primary classifiers make decisions on different types of data, the results of the base classifiers will be fused using ensemble learning.

This embodiment uses an expanded two-stage stacking algorithm to deal with the problem of data imbalance, including two aspects: First, changing the input attribute representation method between levels. The traditional stacking algorithm usually uses the output results of N different primary learners for the same sample data as the characteristic elements of an input vector of the meta-learner, resulting in the continuous growth of the meta-layer feature dimension when the number of primary learners increases, thereby extending the running time of the algorithm. This embodiment uses the output class posterior probabilities of N different primary learners for the same sample as N fixed-dimensional input vectors of the meta-layer classifier, which can not only avoid the expansion of the training data dimension of the meta-layer classifier as the increase of primary classifiers, but also increase the sample content of the training data of the meta-layer learner, thereby saving running time while avoiding data sparsity problems caused by excessive data dimensions.

The second aspect is to increase the number of layers of the stacking algorithm; this embodiment adds a primary layer on the basis of the stacking algorithm, that is, to expand the traditional stacking two-layer algorithm into a three-layer algorithm; because this embodiment has taken into account the purpose of saving running time by changing the input attribute representation method between the levels, the running time of the stacking algorithm after adding the primary layer is not longer than the running time of the traditional stacking algorithm using the same number of primary learners. In addition, when another primary layer is added, it will help to further improve the generalization ability of ensemble learning. The expanded two-stage stacking algorithm of this embodiment increases the number of algorithm layers on the basis of the traditional stacking algorithm, which can not only ensure the accuracy of the diagnostic model estimation results, but also ensure its reliability in extrapolation. It can more effectively solve the challenges brought by unbalanced data sets and improve the accuracy and robustness of the model in predicting minority categories.

In the parameter learning process, the EKM (Ensemble K-modes) method is used to cluster the main category samples; the EKM method can effectively divide the main category samples into different clusters. After clustering, two different data combination strategies are adopted: (1) cluster the majority class samples into K clusters, and then form a new data set s ₁ with the minority class samples; (2) cluster the majority class samples into K clusters, and then divide each cluster into K parts, and select one part from each cluster to form a new balanced sample s ₂ together with the minority class samples.

When carrying out ensemble learning, this embodiment first selects several corresponding primary learners according to different data types. As shown in Figure 3, the primary learners for image data include convolutional neural network CNN, full convolutional network FCN and deep Boltzmann machine DBM, etc.; the primary learners for text data include recurrent neural network RNN, long short-term memory network LSTM and gated recurrent unit GRU, etc.; and the primary learners for numerical data include logistic regression, support vector machine SVM and naive Bayes, etc.

Then, the optimal parameters of the model are selected through grid search and 5-fold cross validation, so as to calculate the conditional probability table at each node in the network. Finally, these conditional probability tables are used for risk assessment and stage assessment for common major diseases. This assessment result can provide an important reference for risk assessment and prevention and treatment of common major diseases.

Verify the effectiveness of the prediction method provided in this embodiment, and evaluate the virtual human prediction model after structure learning and parameter learning; this embodiment refers to the evaluation of the prediction results of the patient's condition. Since the patient's actual condition can be obtained through medical records and follow-up, it is only necessary to compare the consistency between the predicted results of the patient's condition and the actual situation.

Evaluation indicators: Many indicators can be selected for the evaluation of early prediction models for major diseases, such as root mean square error RMSE, mean absolute error MAE and mean absolute percentage error MAPE for quantitative variables, as well as error rate and ROC analysis for categorical variables. Since these indicators are relatively conventional, they will not be repeated. However, it should be noted that the traditional ROC curve is only applicable to the analysis of binary outcomes (such as death), while the outcomes of this embodiment have more than two types of health states (such as extremely low risk, low risk, medium risk, high risk and extremely high risk, etc.). The traditional method for processing multi-class outcome variables is to reclassify the variables according to binary outcomes. However, this approach may bring serious bias to the estimation of classification prediction results. Therefore, this embodiment adopts a high-dimensional ROC analysis method to construct different dimensions of a high-dimensional ROC surface by defining the correct classification rate CCR of different categories. Specifically, the correct classification rate of each dimensional category is used as the coordinate axis to form a coordinate system; and the coordinate positions (i.e., working points) corresponding to different critical value combinations are marked in the coordinate system, and then the high-dimensional ROC surface is formed by connecting the points. Similar to the two-dimensional ROC curve, in the high-dimensional ROC analysis, the volume under the high-dimensional ROC surface VUS is used to measure the accurate discrimination ability of the treatment selection marker for all subjects. The probabilistic statistical meaning of VUS is "after taking an individual from each population to form a new sample, the probability that each individual in the sample is correctly classified into the group to which it actually belongs." This embodiment uses non-parametric and semi-parametric methods to estimate VUS, and their accuracy and precision have been verified in previous studies.

The EKM method is used to cluster the data set. The similarity between the new sample and the training set is mad, and the similarity between the new sample and the correctly classified sample in the training set is mac. That is, it contains four different clustering sample data sets, namely s1 mad, s1 mac, s2mad, and s2 mac. The prediction probability is corrected and the accuracy, sensitivity, specificity and AUC are calculated after passing through four primary learners (LR, C4.5, SVM, KNN), as shown in Figures 4 to 7.

(1) Accuracy: As shown in Figure 4, the performance shows a trend of increasing with the increase of IR. Overall, the performance of s1 mad and s1 mac is better than that of s2 mad and s2 mac.

(2) Sensitivity: As shown in Figure 5, the performance trends are similar. When the IR is small (2, 4), s1 mad and s1 mac perform better than s2 mad and s2 mac, but when the IR is large (16, 32), the results are opposite, and s2 mad and s2 mac perform better than s1 mad and s1 mac.

(3) AUC: As shown in Figure 6, similar to the performance in accuracy and specificity, the performance of the four methods on the four learners shows an increasing trend with the increase of IR. Overall, the performance of s1 mad and s1 mac is better than that of s2 mad and s2 mac.

(4) Specificity: As shown in Figure 7, the performance shows a trend of increasing with the increase of IR, which is similar to the performance in accuracy. Overall, except when IR=2, the performance of s1 mad and s1 mac is better than that of s2 mad and s2 mac on the four classifiers.

The above is only a preferred implementation of the present invention, but the protection scope of the present invention is not limited thereto, and any modification and replacement based on the technical solution and inventive concept provided by the present invention should be included in the protection scope of the present invention.

Claims (6)

1. The method for early prediction of common major diseases based on virtual human simulation is characterized by comprising the following steps of: constructing a virtual human prediction model based on a dynamic Bayesian network through structure learning and parameter learning, and carrying out early prediction on common major diseases through the virtual human prediction model;

The structure learning is used for screening out factors related to risk prediction from potential factors, and determining the topological structure of a dynamic Bayesian network to construct a virtual human prediction model;

The structure learning is to firstly convert knowledge into normalized fuzzy membership degree by utilizing a fuzzy theory, determine the causal relationship between two variables by the membership degree, and construct a plurality of initial virtual human prediction models by taking the relationship between the two variables as a strong limiting condition or a weak limiting condition for searching the Bayesian network structure space; then, an optimal virtual person prediction model is explored from a plurality of initial virtual person prediction models by using a partition MCMC method;

The parameter learning is used for determining probability distribution of each variable in the optimal virtual human prediction model under the condition of giving a father node set, estimating a conditional probability relation among all nodes in the optimal virtual human prediction model by using the existing data set, and distributing probability parameters for each node; the parameter learning takes the stage of common major diseases as a classification ending index of risk assessment, clusters the important diseases through an EKM method, and predicts the result through an amplification type two-stage stacking algorithm; finally, determining optimal parameters of the virtual person prediction model through a grid search method and 5-fold cross verification, so as to construct a final virtual person prediction model;

the amplification type two-stage stacking algorithm takes the posterior probabilities of the output classes of N different primary learners on the same data set as the input vectors of N fixed dimensions of the element layer classifier respectively, and adds one primary layer on the basis of the stacking algorithm.

2. The method for early prediction of common major diseases based on virtual human simulation according to claim 1, wherein the method comprises the following steps: determining the causal relationship between two variables through membership: when the membership is 0, there is no causal relationship between the two variables, when the membership is 1, there is causal relationship between the two variables, and when the membership is between 0 and 1, the causal relationship between the two variables cannot be determined.

3. The method for early prediction of common major diseases based on virtual human simulation according to claim 2, wherein the method comprises the following steps: the strong constraint is the causal relationship of two variables when the membership degree is 0 and the membership degree is 1, and the weak constraint is the causal relationship of two variables when the membership degree is between 0 and 1.

4. The method for early prediction of common major diseases based on virtual human simulation according to claim 1, wherein the method comprises the following steps: the partition MCMC method comprises the following steps:

Dividing all variables in an initial topological structure of the Bayesian network into m zones according to partition requirements and partition rules, numbering the m zones in sequence, setting the variable number of the ith zone (i=1, 2, …, m) as k _i, and setting the variable number Number of variables in each partitionThe specific variable in each partition is pi _λ, the partition Λ= (λ, pi _λ) is marked, and the bayesian network structure under the partition Λ is marked as; Given data D, the posterior distribution of the labeled partition Λ is proportional to the total score obtained by combining the scores of each node X _i and its parent Pa _i in the Bayesian network structureTotal scoreDetermining an optimal network structure in the Bayesian network structure space according to the marked partition with the largest total score for the equivalence of the marked partition space and the Bayesian network structure space;

In each iteration, the current marker partition is Λ, the proposed marker partition is Λ ^*, and the acceptance probability is Wherein, the method comprises the steps of, wherein,To mark partitionsIs partitioned by a markerOne partition is split into two partitions or two adjacent partitions are merged into one partition.

5. The method for early prediction of common major diseases based on virtual human simulation according to claim 4, wherein the method comprises the following steps: the partition requirement comprises that no arrow connection exists between variables in the same region, no father node exists in the variables in the 1 st region, and at least one father node from the previous region is needed in each variable of each region except the 1 st region;

the partitioning rule is determined by a strong constraint, and a random simulation mode is adopted for partitioning variables which are not explicitly specified in the strong constraint.

6. The method for early prediction of common major diseases based on virtual human simulation according to claim 1, wherein the method comprises the following steps: selecting different primary learners according to different types of data sets, wherein the primary learners for image data comprise a convolutional neural network CNN, a full convolutional network FCN and a deep Boltzmann machine DBM; the primary learner for text data includes a cyclic neural network RNN, a long-short-term memory network LSTM and a gating cyclic unit GRU; the primary learner for the numerical data includes logistic regression, support vector machines SVM, and naive bayes.