Causal discovery method for diabetes risk factors based on improved functional causal likelihood

By improving the functional causal likelihood method, an IFCL model for the risk factors of diabetes was established and corrected, which solved the redundant and error problems in causal relationship discovery in the existing technology, achieved more accurate causal structure mining, and improved diabetes Discovery performance of causal relationships of risk factors.

CN112233802BActive Publication Date: 2024-11-19LINGNAN NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202010913729.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-09-03
Publication Date
2024-11-19
Estimated Expiration
2040-09-03

AI Technical Summary

Technical Problem

The existing methods for finding causal relationships of diabetes risk factors have problems with redundant and wrong sides, which is difficult to accurately reflect the causal relationship. Moreover, the FCL algorithm is not effective in discovering causal relationships of diabetes risk factors.

Method used

Using the method of improving the causal likelihood of the function, the FCL model was established by obtaining the joint distribution of the diabetes risk factor variables, and the IFCL model was obtained by adjusting the threshold correction model, and the causal relationship of the risk factors was discovered using the IFCL model to reduce the redundant edges and Error edge.

Benefits of technology

The optimized causal structure of diabetes risk factors has been generated, reducing redundant and error sides, improving the mining performance of the causal discovery algorithm, and able to more accurately reflect the causal relationship.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112233802B_ABST
    Figure CN112233802B_ABST
Patent Text Reader

Abstract

The present invention discloses a causal discovery method for diabetes risk factors based on improved function causal likelihood, which belongs to the field of medical information technology; the method includes obtaining the joint distribution of a subset of diabetes risk factor variables; calculating the log likelihood of observed data according to the joint distribution and the causal structure; converting the log likelihood of the observed data into the log likelihood of the observed data noise, and then establishing a diabetes risk factor FCL model; correcting the diabetes risk factor FCL model by adjusting a threshold, obtaining a diabetes risk factor IFCL model, and using the diabetes risk factor IFCL model to discover the causal relationship of risk factors. The present application introduces an adjustment threshold, constructs a diabetes risk factor IFCL model, uses the diabetes risk factor IFCL model to discover the causal relationship of risk factors, reduces redundant edges and erroneous edges of the diabetes risk factor causal structure, and then generates an optimized diabetes risk factor causal structure.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of medical information technology, and in particular to a causal discovery method for diabetes risk factors based on improved function causal likelihood. Background Art

[0002] The number of patients with diabetes has increased year by year, and it has become the third largest disease threatening human health after cardiovascular and cerebrovascular diseases and malignant tumors. Analyzing the relationship between various risk factors and the relationship between risk factors and diabetes is the key to revealing the pathology of diabetes and is also the premise for the prevention and treatment of diabetes.

[0003] At present, domestic and foreign research mainly focuses on the analysis of diabetes risk factors and the construction of diabetes prediction models. The analysis of diabetes risk factors mainly includes two aspects: the discovery of new risk factors and the analysis of risk factor correlation. Exploring new risk factors helps to discover potential factors for diabetes and understand the causes of diabetes to assist in prevention; at the same time, the onset of diabetes is caused by multiple factors, and the analysis of the relationship between risk factors is also particularly important and has practical clinical significance. Risk factor correlation studies can discover unrecognized physiological and pathological phenomena of diabetes and provide a certain degree of theoretical support for revealing pathology. Diabetes prediction models are helpful for diabetes prevention and early diagnosis. However, the current research on the relationship between risk factors can only reflect correlation, not directly reflect causal relationships; diabetes prediction models are essentially statistical correlation models and cannot reflect causal relationships. Therefore, it is urgent to carry out causal relationship research on diabetes risk factors, and then reveal the pathological and physiological causal relationships of diabetes, which has important theoretical significance and clinical guidance for the prevention and treatment of diabetes.

[0004] Randomized controlled experiments are traditional methods for causal relationship discovery, but they require a lot of intervention in the experimental group, which is not only costly but also may violate ethics. Causal relationship discovery methods based on observational data can avoid the above problems, but the noise in the data will affect the effect of causal relationship discovery. Under significant noise conditions, the FCL algorithm [Reference: Ruichu Cai, Jie Qiao, Zhenjie Zhang, Zhifeng Hao. SELF: Structural Equational Embedded Likelihood Framework for Causal Discovery. AAAI. 2018.] can effectively discover causal relationships. However, this method will retain more redundant edges and erroneous edges in the problem of causal relationship discovery of diabetes risk factors. Summary of the invention

[0005] The causal structure output by the FCL algorithm has the problem of many redundant edges and erroneous edges. The present invention provides a causal discovery method for diabetes risk factors based on improved function causal likelihood, which improves the mining performance of the diabetes risk factor causal discovery algorithm and generates an optimized causal structure of diabetes risk factors.

[0006] To achieve the above object, the technical solution of the present invention is: a causal discovery method for diabetes risk factors based on improved function causal likelihood, comprising:

[0007] Obtain the joint distribution of a subset of diabetes risk factor variables;

[0008] Calculating the log-likelihood of the observed data based on the joint distribution and the causal structure;

[0009] Converting the log-likelihood of the observed data into the log-likelihood of the observed data noise, and then establishing a FCL model of diabetes risk factors;

[0010] The diabetes risk factor FCL model is modified by adjusting the threshold to obtain the diabetes risk factor IFCL model, and the causal relationship of the risk factors is discovered using the diabetes risk factor IFCL model.

[0011] Furthermore, the joint distribution of the subset of diabetes risk factor variables is obtained, including:

[0012] Let {X1,X2,···,X N} is a set of diabetes risk factor variables, where N is the number of diabetes risk factor variables; let G be a subset X = {X1, X2, ···, X n}, n is the number of diabetes risk factor variables with causal relationship; P(X i = x) is the variable X i = the probability of x, where x is the corresponding observed value; For variable X i In all its dependent variables The probability under the condition, where 1≤i≤n≤N, i is the subscript of the diabetes risk factor variable with causal relationship; assuming that the causal structure G satisfies the causal Markov assumption and the causal fidelity assumption, then its joint distribution P(X) is expressed as:

[0013]

[0014] in For variable X i all dependent variables.

[0015] Furthermore, the log-likelihood of the observed data according to the joint distribution and the causal structure specifically includes:

[0016] Given risk factor observation data in is an n-dimensional vector, that is 1≤j≤m; let Indicated as containing The log-likelihood of the observed data is:

[0017]

[0018] Furthermore, the logarithmic likelihood of the observed data is converted into the logarithmic likelihood of the observed data noise, and then the FCL model of diabetes risk factors is established, which specifically includes:

[0019] Assume E i Represents X i The corresponding random noise variable adopts the additive noise model As a causal relationship generating mechanism, F i Yes X i The causal function of the random noise variable E i Independent of the dependent variable Therefore, the following equation is obtained:

[0020]

[0021] From equations (2) and (3), we can see that the log-likelihood of the observed data is equal to the log-likelihood of the observed data noise. Let S =<G,F> is its causal structure, F is the set of causal functions, then the log-likelihood of the observed data noise is:

[0022]

[0023] O is the observed data set. When the data set is limited, equation (4) is regularized to avoid generating too many redundant causal edges. The regularized log-likelihood after adding the Bayesian information criterion is:

[0024]

[0025] Formula (5) is the FCL model of diabetes risk factors, where d i To estimate X i The coefficients used; the causal structure is obtained by maximizing formula (5), maxL B (S; O) = max G sup F L B (<G,F> ; O), which is a causal discovery method based on the FCL model of diabetes risk factors, which includes two processes: ① fitting and optimizing the causal function sup F L B(<G,F> ; O), get the initial causal structure; ② use the hill climbing method to select the causal structure with the largest logarithmic likelihood max G L B (<G,F> ; O), in the iterative update process, X i The local update rule is:

[0026]

[0027] Further, the diabetes risk factor FCL model is modified by adjusting the threshold to obtain the diabetes risk factor IFCL model, and the diabetes risk factor IFCL model is used to find the causal relationship of the risk factors, specifically including: the logarithmic likelihood of the causal structure of the diabetes risk factor after iteration is Since the hill climbing method selects the causal structure with the largest log-likelihood of the target, the termination condition is Where L B (S; O) is the FCL model of the initial causal structure, which leads to the output of the diabetes risk factor causal structure with more redundant edges or erroneous edges; therefore, the adjustment threshold is introduced in formula (5) for correction, and the corrected model is:

[0028]

[0029] Formula (7) is the improved IFCL model of diabetes risk factors, where α is the adjustment threshold; in the hill climbing method, formula (6) is still the X i The local update rule, the iteration termination condition becomes The log-likelihood of nodes that are not updated during the iteration is:

[0030]

[0031] The iterated log-likelihood of the causal structure of diabetes risk factors is expressed as:

[0032]

[0033] Among them, α′ is the total threshold after iteration; from formula (7), we can see that the total threshold of the initial diabetes risk factor IFCL model is nα, which is regarded as each causal node log likelihood has a threshold α, that is, After each iteration using formula (6), the updated node log-likelihood will decrease by a threshold value α, which will cause the total threshold after iteration to be less than the total threshold of the initial diabetes risk factor IFCL model, that is, α′<nα; therefore, during the iteration process, a causal node with a larger log-likelihood must be searched to meet the iteration termination condition. This is the fundamental reason why the diabetes risk factor causal discovery method based on the IFCL model can output a more optimized causal structure.

[0034] The above-mentioned diabetes risk factor causal discovery method based on improved function causal likelihood is specifically implemented as follows:

[0035] Step 1: Enter observational data on diabetes risk factors Preprocess and standardize observational data;

[0036] Step 2: First, use the regression method to estimate the causal function F corresponding to the causal edge i ; Secondly, calculate the noise norm obtained by regression; then use kernel density estimation to approximate the distribution of noise to optimize the causal function F i ; Finally, the initial causal structure G is obtained;

[0037] Step 3: Initialize the noise log likelihood using equation (7) And let the variable

[0038] Step 4: Use the hill climbing method to search for the optimal causal structure; in each iteration, only one causal edge in the causal structure G is added, deleted, or the causal direction is flipped; update the causal function F i and causal structure, and store the updated causal structure in G * ;

[0039] Step 5: Compare G * With G, the local log-likelihood of the changed nodes is updated to obtain L′ Bi , where the update formula is (6); then all updated log-likelihoods With the unupdated log-likelihood Add, and

[0040] Step 6: Judgement With L B The size of but G=G * And execute Step 7; otherwise execute Step 4;

[0041] Step 7: Output the maximum log-likelihood And the corresponding optimal causal structure G.

[0042] By adopting the above technical scheme, the present invention can achieve the following technical effects: introducing an adjustment threshold, constructing an IFCL model of diabetes risk factors, using the IFCL model of diabetes risk factors to discover the causal relationship of risk factors, reducing redundant edges and erroneous edges of the causal structure of diabetes risk factors, and then generating an optimized causal structure of diabetes risk factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flow chart for discovering the causal relationship of risk factors using the IFCL model of diabetes risk factors;

[0044] Figure 2 Scatter plots, bar graphs, and correlation coefficients for the variables with a sample size of 768;

[0045] Figure 3 Scatter plots, bar graphs, and correlation coefficients for variables with a sample size of 2000;

[0046] Figure 4 This is a causal structure diagram based on the FCL model of diabetes risk factors, with a sample size of 768;

[0047] Figure 5 This is a causal structure diagram based on the FCL model of diabetes risk factors, with a sample size of 2000;

[0048] Figure 6 This is the causal structure diagram based on the IFCL model of diabetes risk factors, with a sample size of 768 and a threshold of α = 0.05-0.06;

[0049] Figure 7 The causal structure based on the IFCL model of diabetes risk factors, with a sample size of 768 and a threshold of α = 0.07-0.14;

[0050] Figure 8 This is the causal structure diagram based on the IFCL model of diabetes risk factors, with a sample size of 768 and a threshold of α = 0.15;

[0051] Fig. 9 This is the causal structure diagram based on the IFCL model of diabetes risk factors, with a sample size of 2000 and a threshold of α = 0.05-0.06;

[0052] Fig.10 This is the causal structure diagram based on the IFCL model of diabetes risk factors, with a sample size of 2000 and a threshold of α = 0.07-0.15;

[0053] Fig.11 This is the causal structure diagram based on the IFCL model of diabetes risk factors, with a sample size of 2000 and a threshold of α = 0.18. DETAILED DESCRIPTION

[0054] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments: this example is used to further describe the present application.

[0055] Example

[0056] In order to verify the feasibility and effectiveness of the present invention, three experiments were conducted, namely, the scatter plot analysis experiment between variables and their correlation coefficients, the causal discovery experiment based on the FCL model of diabetes risk factors, and the causal discovery experiment based on the IFCL model of diabetes risk factors. The experimental data are diabetes datasets with a sample size of 768 [https: / / www.kaggle.com / uciml / pima-indians-diabetes-database] and 2000 [https: / / www.kaggle.com / uciml / pima-indians-diabetes-database] on the Kaggle platform, which are from the National Institute of Diabetes, Digestive and Kidney Diseases of the United States and the Frankfurt Hospital of Germany, respectively. The subjects of the dataset are all over 21 years old, and the dataset includes 9 variables, namely, the number of pregnancies, 2-hour plasma glucose concentration (referred to as blood sugar) in the oral glucose tolerance test, diastolic blood pressure (mmHg), triceps skin fold thickness (mm), 2-hour serum insulin (muU / ml, referred to as insulin), body mass index (BMI), diabetes ancestry function, age and diabetes diagnosis binary variables, where the diabetes ancestry function contains the subject's family diabetes genetic information. The binary variable of diabetes diagnosis was eliminated, and only 8 variables were analyzed for causal relationship. In order to retain the data information to the greatest extent, the mean interpolation method was used to fill the missing values ​​in the data set. The original data was Z-score standardized, and the mean was used to replace the outliers.

[0057] (1) Scatter plots and correlation coefficient analysis experiments between variables

[0058] In order to observe the causal relationship between the eight variables, we first analyze the scatter plots of the two variables and their correlation coefficients. The scatter plots of the variables with a sample size of 768 and a sample size of 2000 are as follows: Figure 2 and 3 As shown. Figure 2 and Figure 3 In the figure, the lower left corner is a variable scatter plot, the image on the diagonal line is a variable bar graph, and the upper right corner is the linear correlation coefficient between the variables. Generally, when the correlation coefficient is between 0 and 0.1, the variables can be considered linearly independent, so these variables are eliminated, and the linear correlation coefficients greater than 0.1 and the significance test results are summarized in Tables 1 and 2. From Tables 1 and 2, it can be seen that all P values ​​are less than 0.01, indicating that there is a linear correlation between the variables.

[0059] Table 1 is a summary table of variable correlation coefficients and P values ​​for a sample size of 768;

[0060]

[0061]

[0062] Table 2 is a summary table of variable correlation coefficients and P values ​​for a sample size of 2000;

[0063]

[0064] (2) Causal discovery experiment based on the FCL model of diabetes risk factors

[0065] To explore whether there is causality between these eight variables, the causal discovery method of the FCL model of diabetes risk factors was used to conduct a causal discovery experiment on data sets with sample sizes of 768 and 2000. The causal structure is as follows Figure 4 and 5 As shown, the corresponding maximum likelihood is summarized in Table 3, where Figure 4-5 The causal structures are named Structure 1-2.

[0066] Table 3 is a summary table of the maximum likelihood of causal structures 1-2;

[0067]

[0068] ① Figure 4 and Figure 5 Similarities. There are 6 pairs of identical causal relationships: 1→7, 7→3, 4→6, 5→2, 2→6, and 6→3, where 1→7 means that the number of pregnancies leads to age changes, 7→3 means that age leads to diastolic blood pressure changes, 4→6 means that triceps skin fold thickness leads to BMI changes, 5→2 means that insulin leads to blood sugar changes, 2→6 means that blood sugar leads to BMI changes, and 6→3 means that BMI leads to diastolic blood pressure changes; in addition, there is no causal relationship between the diabetes ancestry function and other variables.

[0069] ② Figure 4 and Figure 5 The difference. Figure 4 There is a causal relationship between 6 and 7. Figure 5 There are two pairs of causal relationships, 7→2 and 4→7, where 6→7 means BMI causes age changes, 7→2 means age causes blood sugar changes, and 4→7 means triceps skin fold thickness causes age changes.

[0070] from Figure 2It can be seen that the correlation coefficient between BMI and age is 0.07, and its P value is 0.072, so BMI and age can be considered to have no linear correlation. Obviously, the causal function obtained by the regression method did not pass the significance test and has no statistical significance. 6→7 can be considered as a false causal relationship. Figure 5 It can be seen that when using 2000 samples for causal relationship discovery, the wrong causal edges such as 6→7 are removed; but at the same time, other causal edges such as 4→7 and 7→2 are also found. At the same time, it can be seen from Table 3 that Figure 5 Compare Figure 4 The maximum likelihood is greater. It can be seen that the size of the sample affects the results of causal relationship discovery. The larger the sample size, the more conducive it is to removing incorrect causal edges and discovering other causal edges; however, the more causal edges are discovered, the more redundant edges of the causal structure may be. Figure 4 and Figure 5 It can be seen that the structure of its causal structure graph is complex and contains many redundant or erroneous edges, and new causal discovery methods need to be studied.

[0071] (3) Causal discovery experiment based on the IFCL model of diabetes risk factors

[0072] The causal discovery method based on the IFCL model of diabetes risk factors proposed in the present invention was used to conduct causal discovery experiments on data sets with sample sizes of 768 and 2000. The flowchart of the method is shown in the figure. Figure 1 The result is closely related to the value of the threshold α. In the experiment, the threshold range is 0.05≤α≤0.18, and the value interval is 0.01. When α<0.05, the generated causal structure graph has more redundant edges, which are not considered here.

[0073] ①Causal structure analysis of a data set with a sample size of 768. When α = 0.05-0.06, the generated causal structure (Structure 3) is as follows Figure 6 As shown, Figure 4 In comparison, the wrong edges in the causal structure, namely 6→7 and 7→3, are removed, and the causal edge 6→3 is reversed to 3→6; when α = 0.07-0.14, the generated causal structure (structure 4) is as follows Figure 7 As shown, Figure 6 In comparison, the causal edge 5→2 is removed; when α = 0.15, the generated causal structure (structure 5) is as follows Figure 8 As shown in the figure, the causal structure is simplified to only two causal edges 1→7 and 4→6, and as α increases, the causal structure graph is no longer simplified.

[0074] ②Causal structure analysis of a data set with a sample size of 2000. When α = 0.05-0.06, the generated causal structure (Structure 6) is as follows Fig. 9 As shown, Figure 5In comparison, the causal edges 7→3 and 7→2 are removed, and the causal edge 6→3 is reversed to 3→6; when α = 0.07-0.15, the generated causal structure (structure 7) is as follows Fig.10 As shown, Fig. 9 In comparison, the causal edge 5→2 is eliminated; when α=0.16-0.17, the optimal causal structure cannot be searched; when α≥0.18, the generated causal structure (structure 8) is as follows Fig.11 As shown in the figure, the causal structure is simplified to only two causal edges 1→7 and 4→7, and as α increases, the causal structure graph no longer changes.

[0075] The maximum likelihoods of causal structures 3-8 are summarized in Table 4. The larger the sample size, the larger the corresponding maximum likelihood of the causal structure.

[0076] Table 4 is a summary table of the maximum likelihood of causal structures 3-8

[0077]

[0078] From the above analysis, we can get the following results. The larger the sample size, the fewer false causal relationships, and other potential causal relationships can be mined at the same time; in the process of causal discovery, it is necessary to add a threshold to reduce redundant edges and false edges. As the threshold α increases, the causal structure of the method of the present invention is continuously simplified, and has the most simplified structure; when α is 0.05 and 0.06, the causal structure has the least redundant edges and retains the most information. α of 0.05 and 0.06 is the best threshold for causal discovery of diabetes risk factors.

[0079] In order to further analyze the rationality of the experimental results, the experimental results are discussed in depth. Figure 6 and Fig. 9 As shown, there are 6 pairs of causal relationships between diabetes risk factors (5→2, 2→6, 4→6, 3→6, 1→7 and 4→7), which are analyzed one by one below.

[0080] ①Causal relationship 5→2, 2→6. This causal relationship is well known to the public. Insulin is the only hormone in the human body that lowers blood sugar and affects blood sugar concentration. If insulin resistance occurs, blood sugar consumption in the body will become abnormal, blood sugar concentration will increase, and diabetes can easily occur. In addition, the causal relationships 5→2 and 2→6 can be combined into 5→2→6. In a normal human body, insulin secretion is normal, blood sugar is metabolized normally, and BMI will remain normal. Diabetic patients have insulin resistance, and blood sugar cannot be taken up and utilized by the liver, muscle and adipose tissue, which will cause the patient to lose weight and reduce BMI. This shows that the causal relationship 5→2→6 is also true.

[0081] ②Causal relationship 4→6. The thickness of the triceps skin folds can reflect the amount of fat in the body. The thicker the triceps skin folds, the more obese and heavier the body weight, which leads to a higher BMI and an increased risk of diabetes. Diabetic patients will lose weight, the thickness of the triceps skin folds will decrease, the weight will decrease, and the BMI will also decrease. Therefore, the causal relationship 4→6 is also established.

[0082] ③Causal relationship 3→6. When using the method in the literature [Ruichu Cai, Jie Qiao, Zhenjie Zhang, Zhifeng Hao. SELF: Structural Equational Embedded Likelihood Framework for Causal Discovery. AAAI. 2018.] to discover causal relationships, the relationship between 3 and 6 is 6→3 (e.g. Figure 4 , Figure 5 As shown in Figure 2, BMI affects diastolic blood pressure. The more obese a person is, the greater his BMI is, and the more likely he is to develop high blood pressure. This is in line with medical common sense and shows that BMI is one of the causes of high blood pressure. However, when the method of the present invention is used to find the causal relationship, the relationship between 3 and 6 is 3→6 (as shown in Figure 2). Figure 6 , Fig. 9 As shown in the figure, the explanation of the causal relationship 3→6 may involve a potential causal factor. Diastolic blood pressure causes the potential factor to change, which causes the BMI to change, thus affecting the risk of diabetes. In addition, diabetic patients are accompanied by hypertension and lose weight. Medically, diabetes is believed to cause patients to lose weight, but it may also be a combination of diabetes and hypertension, which in turn affects BMI. Therefore, the causal relationship 3→6 may exist, but it is not well known.

[0083] ④Causal relationship 1→7. It indicates that the number of pregnancies will lead to age changes. The literature [Ryan CP, Geoffrey HM, Lee NR, et al. Reproduction predicts shorter telomeres and epigenetic age acceleration among young adult women [J]. Scientific Reports, 2018, 8 (1): 11100.] found that the more pregnancies there are, the older the physiological age is, that is, the cells will age faster and the probability of developing certain diseases will also increase. Therefore, the causal mechanism predicted by the causal relationship 1→7 may be: as the number of pregnancies increases, the pancreatic β cells will age faster, insulin resistance will occur more easily, and the risk of diabetes will increase.

[0084] ⑤Causal relationship 4→7. It indicates that the thickness of triceps skin folds will lead to age changes. The thickness of triceps skin folds can reflect a person's nutritional status. The causal mechanism predicted by causal relationship 4→7 may be: if the thickness of triceps skin folds is lower than the standard value, malnutrition will occur, affecting physiological age and aging of pancreatic β cells, leading to insulin resistance and increasing the probability of diabetes; if the thickness of triceps skin folds exceeds the standard value, it means obesity. Excessive glucose in the body causes pancreatic β cells to work overload for a long time, which is easy to cause pancreatic aging and functional damage, and the risk of diabetes increases.

[0085] In summary, the causal relationships confirmed by the method for causal discovery of diabetes risk factors proposed in the present invention are: 5→2→6 and 4→6; the causal relationships that need to be further verified are: 3→6, 1→7 and 4→7, which shows that the method of the present invention has great potential in the causal discovery of diabetes risk factors, helps to further reveal the causal relationship between risk influencing factors, and provides a certain reference for the causal relationship mining of diabetes risk factors.

[0086] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A causal discovery method for diabetes risk factors based on improved function causal likelihood, characterized in that: include: Get the joint distribution of a subset of diabetes risk factor variables, including: Let {X1,X2,···,X N } is a set of diabetes risk factor variables, where N is the number of diabetes risk factor variables; let G be a subset X = {X1, X2, ···, X n }, n is the number of diabetes risk factor variables with causal relationship; P(X i =x) is the variable X i = the probability of x, where x is the corresponding observed value; For variable X i In all its dependent variables The probability under the condition, where 1≤i≤n≤N, i is the subscript of the diabetes risk factor variable with causal relationship; assuming that the causal structure G satisfies the causal Markov assumption and the causal fidelity assumption, then its joint distribution P(X) is expressed as: in For variable X i All dependent variables; The log-likelihood of the observed data is calculated based on the joint distribution and causal structure, specifically including: Given risk factor observation data in is an n-dimensional vector, that is make Indicated as containing The log-likelihood of the observed data is: The logarithmic likelihood of the observed data is converted into the logarithmic likelihood of the observed data noise, and then the FCL model of diabetes risk factors is established, which specifically includes: Assume E i Represents X i The corresponding random noise variable adopts the additive noise model As a causal relationship generating mechanism, F i Yes X i The causal function of the random noise variable E i Independent of the dependent variable Therefore, the following equation is obtained: in, Indicated as containing The observed value of For X i All dependent variables; From equations (2) and (3), we can see that the log-likelihood of the observed data is equal to the log-likelihood of the observed data noise. Let S =<G,F> is its causal structure, F is the set of causal functions, then the log-likelihood of the observed data noise is: O is the observed data set, and the regularized log-likelihood after adding the Bayesian information criterion is: Formula (5) is the FCL model of diabetes risk factors, where d i To estimate X i The coefficients used; the causal structure is obtained by maximizing formula (5), maxL B (S; O) = max G sup F L B (<G,F> ; O), which is a causal discovery method based on the FCL model of diabetes risk factors, which includes two processes: ① fitting and optimizing the causal function sup F L B (<G,F> ; O), get the initial causal structure; ② use the hill climbing method to select the causal structure with the largest logarithmic likelihood max G L B (<G,F> ; O), in the iterative update process, X i The local update rule is: The FCL model of diabetes risk factors is modified by adjusting the threshold to obtain the IFCL model of diabetes risk factors, and the IFCL model of diabetes risk factors is used to find the causal relationship of risk factors, which specifically includes: introducing the adjustment threshold into formula (5) for modification, and the modified model is: Formula (7) is the improved IFCL model of diabetes risk factors, where α is the adjustment threshold; in the hill climbing method, formula (6) is still the X i The local update rule, the iteration termination condition becomes The log-likelihood of nodes that are not updated during the iteration is: The iterated log-likelihood of the causal structure of diabetes risk factors is expressed as: Among them, α′ is the total threshold after iteration; from formula (7), we can see that the total threshold of the initial diabetes risk factor IFCL model is nα, which is regarded as each causal node log likelihood has a threshold α, that is, After each iteration using formula (6), the updated node log-likelihood will decrease by a threshold value α, which will cause the total threshold after iteration to be less than the total threshold of the initial diabetes risk factor IFCL model, that is, α′<nα; therefore, during the iteration process, a causal node with a larger log-likelihood must be searched to meet the iteration termination condition. This is the fundamental reason why the diabetes risk factor causal discovery method based on the IFCL model can output a more optimized causal structure.