An integrated adaptive neurofuzzy system for diabetes analysis

Abstract

This invention discloses an adaptive neurofuzzy system for diabetes analysis and integration, comprising the following steps: S1: inputting the dataset into a CIR-ANFIS model, where the model first standardizes the dataset; S2: calculating the causal coefficient corresponding to the prediction result for each feature using causal inference, and using this coefficient as the feature weight; S3: randomly selecting multiple different subsets, constructing an adaptive fuzzy inference system for each subset, and obtaining the prediction result of each inference system; S4: obtaining the prediction result of the entire model through ensemble learning. This invention relates to the field of computer algorithm technology. The beneficial effect of this invention is that by introducing causal inference into feature selection and using causal coefficients as weights to randomly generate data subsets for training, the model training cost is reduced, and better model performance is achieved.

Description

Technical Field

[0001] This invention relates to the field of computer algorithms, and in particular to an integrated adaptive neurofuzzy system for diabetes analysis. Background Technology

[0002] Adaptive Neural Fuzzy Systems (ANFIS) have become a popular disease prediction tool due to their interpretability. However, the structure of ANFIS dictates that the number of parameters increases exponentially with the number of input features. Simply reducing the number of features will decrease the accuracy of the resulting model, while adding more features will increase the cost of training the model and reduce the interpretability of ANFIS, severely limiting its application in situations requiring a high level of interpretability, especially disease prediction. In this paper, we propose an Integrated Stochastic Causal Inference-Adaptive Neural Fuzzy System (CIR-ANFIS) model. It uses causal inference to determine the causal relationship between features and outcomes, and uses this randomly generated subset to train different ANFIS models. Then, various bagging strategies are developed based on the unique characteristics of ANFIS for comparison. Comparative experiments show that the proposed model significantly outperforms traditional ANFIS models, other feature selection techniques, and feature selection using causal inference on other datasets for diabetes and other disease predictions. The CIR-ANFIS model described in this study can also be used to classify binary disease prediction scenarios, particularly those based on laboratory indicators.

[0003] To build a CIR-ANFIS model for predicting diabetes and apply it to other medical datasets, a novel approach was taken: using causal inference (CI) for feature selection and ensemble learning to reduce underfitting and overfitting of individual ANFIS models. Ensemble learning and causal inference (CI) will be discussed in the following sections.

[0004] Technical defects:

[0005] First, while ensemble learning is more accurate than using a single model, it is less efficient due to the more complex training process. This is especially true when the number of features is high, as the number of sub-models increases significantly, and the training time for CIR-ANFIS also increases.

[0006] Second, the ensemble learning model used was a semi-persuasive model. To achieve better results, more ensemble learning methods will be explored in the future. Furthermore, this project used a technique of summing and averaging each sub-data model to obtain the results, with each model having equal weights; further exploration of better combination methods is needed. Summary of the Invention

[0007] The purpose of this invention is to solve the above problems by designing an adaptive neurofuzzy system for diabetes analysis and integration.

[0008] The technical solution of the present invention to achieve the above objectives is an adaptive neurofuzzy system applied to diabetes analysis and integration, comprising step S1: inputting the dataset into the CIR-ANFIS model, wherein the model first standardizes the dataset;

[0009] Step S2: Calculate the causal coefficient of each feature corresponding to the prediction result using causal reasoning, and use this coefficient as the feature weight;

[0010] Step S3: Randomly select multiple different subsets, construct an adaptive fuzzy inference system for each different subset, and obtain the prediction results of each inference system;

[0011] Step S4: Obtain the prediction results of the entire model through ensemble learning.

[0012] In step S1, the input dataset needs to be standardized before training the CIR-ANFIS model.

[0013] In this section, all features are converted to a form with a mean of 0 and a variance of 1. The steps are as follows:

[0014] (1) For each feature, calculate its mean and then subtract its mean;

[0015] (2) Calculate the variance of each feature, and then divide the result of (1) by the variance.

[0016] The weight of each feature in step S2 may be determined by its causal relationship with the outcome; this indication is called causal effect. In statistics, causal inference is the experimental study that determines causal relationships. The following are the four steps of causal inference:

[0017] Step 1: Model the problem based on assumptions;

[0018] Since causal models are developed using problem assumptions, the prior knowledge used in their construction can be used to incorporate past information and create a graph.

[0019] Without prior knowledge of the dataset, it is assumed that there is no causal relationship between features and all features are affected by unobserved confounding factors. When determining the causal relationship between a specific feature and the outcome, it is assumed that if the feature is above the mean and the outcome is positive, then there is a causal relationship. To this end, normalized features are converted into 0-1 variables and the causal relationship is calculated using these variables. If there is prior knowledge of the dataset, this knowledge can be used to create a more refined causal inference graph, which helps to obtain more accurate results.

[0020] Step 2: Determine the causal relationship (causal estimation);

[0021] The definition of causal analysis indicates that changes in an intervention will affect the outcome when other variables remain constant. After determining the causal effect of the causal model, it can be estimated based on previous assumptions. In this study, the backdoor criterion was mainly used.

[0022] Step 3: Use statistical methods to estimate;

[0023] In this step, the causal relationship between each feature and outcome can be calculated. Many techniques, such as propensity score matching, propensity score stratification, propensity score-based inverse probability weighting, linear regression, generalized linear models (such as logistic regression), instrumental variables, and point regression, have been used to determine statistical causal effects.

[0024] The causal effect ranges between -1 and 1, where a negative value indicates a negative correlation between the feature and the outcome.

[0025] Step 4: Use multiple robustness testing methods to verify the credibility of the causal effect;

[0026] Although statistical estimations of the data were used in Section 3 to calculate causality, the causality itself is based on prior assumptions rather than the data. Therefore, several robustness tests were employed to demonstrate the prior assumptions. The three robustness tests used in this study are as follows:

[0027] 1) Random confounding factor;

[0028] In this test, independent random variables are introduced as common influencing factors of the dataset. If the assumptions are accurate, the estimation results will not change naturally.

[0029] 2) Data subset validation;

[0030] In this test, a randomly selected subset is used instead of the provided dataset, and if the assumptions are true, the estimation results will not change significantly.

[0031] 3) Placebo intervention;

[0032] In this test, independent random variables are used instead of the actual intervention variables; if the assumptions are valid, the estimated results will be close to zero.

[0033] The causal effect can be considered as the influence of each feature on the classification. Features with a high level of causal influence may contribute more than features with a low level of causal influence. Using the causal effect as the feature weight, a random subset is generated through the following operation.

[0034] In the CIR-ANFIS model, using There are 4 sub-models, where n is the number of input features. Each sub-model has four input features. For the 2CIR-ANFIS model, twice the number of input features is used. The following is an algorithm for selecting the input features for each sub-model:

[0035] (1) For each feature, the weight score The definition is as follows:

[0036] ;

[0037] in It is the absolute value of the causal relationship between the feature and the result. It is a random number between 0 and 1.

[0038] (2) Select the four features with the highest weight scores.

[0039] (3) If this is not the first sub-model, compare the selection result with the previous result. If all four features match, the selection result is invalid; return to step 1 and perform the calculation again; otherwise, continue to calculate each subsequent sub-model until all sub-data has been calculated.

[0040] The above process completes the steps for preparing the sub-dataset for sub-model training. To demonstrate that this method can utilize more features, it is compared with the traditional method of simply selecting four features;

[0041] In CIR-ANFIS, for those with The probability of a dataset with 1 feature not being selected by the submodel. It can be approximated as

[0042] ;

[0043] The probability of not being selected by all sub-models for

[0044] ;

[0045] Continue the analysis under various conditions;

[0046] (1) It is significantly higher than most other common traits, ranking in the top 4. This characteristic is likely to be included in the subset multiple times, which means that it can contribute to many sub-models. Obviously, it makes a great contribution to the entire CIR-ANFIS model.

[0047] (2) If the ranking is below or above the Top-4, but the difference is not significant, then... The value ranges from 0.1 to 0.2. The value ranges from 0.4 to 0.6, which indicates that this feature can still be incorporated into some subsets of the dataset. In other words, the CIR-ANFIS model allows for the inclusion of more features rather than just selecting a limited number of features.

[0048] (3) Significantly lower than the Top-4, for example, one-twentieth or even one-hundredth of the Top-4. In this case, Below 0.05, A value higher than 0.8 means that due to the low causal relationship between features and results, features may sometimes only be included in one or a few subsets, or even not in any subsets at all.

[0049] Based on the above calculations and analysis, CIR-ANFIS can use causal inference to create subset datasets that reflect the relative importance of various features, with the aim of using more data from more features and improving the performance of the overall model.

[0050] The ANFIS mentioned is a classic TSK fuzzy model, which adopts a Takagi-Sugeno-Kang type first-order rule, defined as follows:

[0051] ;

[0052] in, It is the i-th feature of the data. It is a fuzzy set, and y is the result of the rule. These are the result coefficients of the TSK rule. The result part in ANFIS is a linear function of the input vector, rather than a fuzzy rule of the Mamdani type. Using a linear function makes ANFIS easier to understand compared to other more complex functions.

[0053] A typical ANFIS consists of 5 layers;

[0054] In the first layer, membership functions are used to blur the input samples. The number of membership functions can be adjusted before training the ANFIS model. Various function shapes are widely used in this layer; in this study, triangular membership functions are used, i.e.

[0055] ;

[0056] Where a, b, and c are the three vertices of the triangle, and a feature has three membership functions to describe the low, medium, and high levels of the feature;

[0057] The first layer can be described as follows:

[0058] ;

[0059] The second layer uses membership degrees to calculate the excitation intensity of samples based on all rules. In this paper, all sub-ANFIS models have four input features, and the model is rule-based. The output of the second layer is the product of the membership degrees of all features.

[0060] ;

[0061] The third layer completes the normalization of the output of the previous layer:

[0062] ;

[0063] It is the normalized excitation intensity;

[0064] The fourth layer calculates the results of all rules, using a linear function to combine the inputs:

[0065] ;

[0066] Finally, the fifth layer completes the deblurring process:

[0067] ;

[0068] For ANFIS, the result of a sample calculation is a floating-point number, but what is needed is a binary 0-1 variable, so rounding can solve this problem.

[0069] The method of combining multiple independent models through ensemble learning to improve generalization performance includes three strategies in current ensemble learning techniques: bagging, boosting, and stacking.

[0070] Unlike most models, the ANFIS model yields twice the expected result, defining a result above 0.5 as positive and below 0.5 as negative. Several bagging strategies will be tested:

[0071] (1) Voting mode

[0072] ;

[0073] ;

[0074] Where m is the number of sub-models, and in CIR-ANFIS, m is the number of features minus 3; res is the prediction of each sub-model. The result is the prediction of the entire model. If more than or equal to 50% of the sub-models correctly predict the positive class result of a sample, the final result is positive.

[0075] (2) Persuasion Model

[0076] ;

[0077] The variables have the same meaning as above. When using this strategy, if a sub-model strongly tends to be positive or negative, a positive or negative class may affect the result of the entire model.

[0078] (3) Persuasion pattern with inhibition function

[0079] ;

[0080] Use continuous functions to reduce the negative impact of extreme values ​​on classification performance, including square root functions and logarithmic functions:

[0081] ;

[0082] ;

[0083] (4) Semi-persuasion model

[0084] ;

[0085] Here, a and b are upper and lower bounds, which limit the ability to "persuade" other sub-models. They can be floating-point numbers or infinity. In this paper, two semi-persuasion methods will be attempted: a is 0 and b is infinity, or a is 0 and b is 1.

[0086] After five steps, the CIR-ANFIS model is obtained.

[0087] An adaptive neurofuzzy system for diabetes analysis and integration was created using the technical solution of the present invention. (1) Causal reasoning method was introduced into feature selection, and a subset of data was randomly generated for training with causal coefficient as weight, which reduced the model training cost and achieved better model performance.

[0088] (2) Unlike conventional ensemble methods, a “semi-persuasive” approach is used to combine the prediction results of all sub-models;

[0089] (3) Use experimental data to demonstrate the performance of the model. Attached Figure Description

[0090] Figure 1 This is a flowchart of a method for an integrated adaptive neurofuzzy system applied to diabetes analysis, as described in this invention. Detailed Implementation

[0091] The present invention will now be described in detail with reference to the accompanying drawings, such as... Figure 1As shown, this patent describes an integrated adaptive neurofuzzy system applied to diabetes analysis. It implements an integrated stochastic causal reasoning-adaptive neurofuzzy system (CIR-ANFIS) model. When a dataset is input into the model, it first standardizes the dataset, then uses causal reasoning to calculate the causal coefficient of each feature corresponding to the prediction result, and uses this coefficient as the feature weight. Multiple different subsets are randomly selected, and an adaptive fuzzy reasoning system is constructed for each subset. The prediction result of each reasoning system is obtained, and finally, the prediction result of the entire model is obtained through ensemble learning.

[0092] The steps are as follows:

[0093] Before training the CIR-ANFIS model, the input dataset needs to be standardized.

[0094] In this section, all features are converted to a form with a mean of 0 and a variance of 1. The steps are as follows:

[0095] (1) For each feature, calculate its mean and then subtract its mean;

[0096] (2) Calculate the variance of each feature, and then divide the result of (1) by the variance.

[0097] Causal reasoning

[0098] The weight of each feature may be determined by its causal relationship with the outcome; this indication is called causal effect. In statistics, causal inference is experimental research that determines causal relationships. Here are the four steps of causal inference:

[0099] Problem modeling based on assumptions

[0100] Since causal models are developed using problem assumptions, they are constructed using prior knowledge. The model can be used to incorporate past information and create a graph.

[0101] Without prior knowledge of the dataset, it is assumed that there are no causal relationships between features, and all features are influenced by unobserved confounding factors. When determining a causal relationship between a specific feature and an outcome, it is assumed that a causal relationship exists if the feature is above the mean and the outcome is positive. To this end, normalized features are converted into 0–1 variables, and causal relationships are calculated using these variables. If prior knowledge of the dataset is available, this knowledge can be used to create a more refined causal inference graph, which helps to obtain more accurate results.

[0102] Determining causal relationships (causal estimation)

[0103] Causal analysis is defined as an intervention that affects the outcome when other variables remain constant. After determining the causal effect of a causal model, estimates can be obtained based on prior assumptions. In this study, the backdoor criterion was primarily used.

[0104] Estimation using statistical methods

[0105] In this step, the causal relationship between each feature and outcome can be calculated. Many techniques, such as propensity score matching, propensity score stratification, propensity score-based inverse probability weighting, linear regression, generalized linear models (such as logistic regression), instrumental variables, and point regression, have been used to determine statistical causal effects.

[0106] The causal effect ranges between -1 and 1, where negative means there is a negative correlation between the feature and the outcome.

[0107] The reliability of the causal effect was verified using a variety of robustness testing methods.

[0108] Although statistical estimations of the data were used in Section 3 to calculate causality, the causal relationships themselves are based on prior assumptions rather than the data. Therefore, several robustness tests were employed to demonstrate the prior assumptions. The three robustness tests used in this study are as follows:

[0109] 1) Random confounding factor

[0110] In this test, independent random variables were introduced as common influencing factors in the dataset. If the assumptions are accurate, the estimation results will not change naturally.

[0111] 2) Data subset validation

[0112] In this test, a randomly selected subset is used instead of the provided dataset; if the assumptions are true, the estimation results will not change significantly.

[0113] 3) Placebo intervention

[0114] In this test, independent random variables are used instead of the actual intervention variables; if the assumptions are valid, the estimated results will be close to zero.

[0115] Generate random subsets

[0116] Causal effect can be considered as the influence of each feature on the classification. Features with high levels of causal influence may contribute more than features with low levels of causal influence. Using causal effect as feature weights, random subsets are generated through the following operation.

[0117] In the CIR-ANFIS model, using There are 2 sub-models, where n is the number of input features, and each sub-model has four input features. For the 2CIR-ANFIS model, twice the number of ( The following is the algorithm for selecting the input features for each sub-model:

[0118] (1) For each feature, the weight score The definition is as follows:

[0119] ;

[0120] in It is the absolute value of the causal relationship between the feature and the result. It is a random number between 0 and 1.

[0121] (2) Select the four features with the highest weight scores.

[0122] (3) If this is not the first sub-model, compare the selection result with the previous result. If all four features match, the selection result is invalid; return to step 1 and perform the calculation again; otherwise, continue to calculate each subsequent sub-model until all sub-data has been calculated.

[0123] The above process completes the steps for preparing the subset dataset for sub-model training. To demonstrate that this method can utilize more features, it is compared with the traditional method of simply selecting four features.

[0124] In CIR-ANFIS, for those with The probability of a dataset with 1 feature not being selected by the submodel. It can be approximated as

[0125] ;

[0126] The probability of not being selected by all sub-models for

[0127] ;

[0128] The discussion will continue in various scenarios.

[0129] (1) It is significantly higher than most other common traits, ranking in the top 4. This characteristic is likely to be included in the subset multiple times, meaning it can contribute to many sub-models, and obviously, it makes a significant contribution to the entire CIR-ANFIS model.

[0130] (2) Below or above the Top-4, but the difference is not significant. In this case, The value ranges from 0.1 to 0.2. The value ranges from 0.4 to 0.6, indicating that this feature can still be incorporated into some subsets of the dataset. In other words, the CIR-ANFIS model allows for the inclusion of more features rather than just selecting a limited number of features.

[0131] (3) Significantly lower than the Top-4, for example, one-twentieth or even one-hundredth of the Top-4. In this case, Below 0.05, A value higher than 0.8 means that due to the low causal relationship between features and results, features may sometimes only be included in one or a few subsets, or even not in any subset at all.

[0132] Based on the above calculations and analysis, CIR-ANFIS can use causal inference to create subset datasets that reflect the relative importance of various features, with the aim of using more data from more features and improving the performance of the overall model.

[0133] Training ANFIS sub-model

[0134] ANFIS is a classic TSK fuzzy model. This model employs a Takagi-Sugeno-Kang type first-order rule, defined as...

[0135] ;

[0136] in, It is the i-th feature of the data. It is a fuzzy set, and y is the result of the rule. These are the result coefficients of the TSK rule. The result part of ANFIS is a linear function of the input vector, rather than a fuzzy rule like Mamdani's. Using a linear function makes ANFIS easier to understand compared to other more complex functions. A typical ANFIS structure is as follows... Figure 1 As shown.

[0137] A typical ANFIS consists of 5 layers.

[0138] Membership functions are used in the first layer to blur the input samples. The number of membership functions can be adjusted before training the ANFIS model. Various function shapes are used extensively in this layer. In this study, triangular membership functions are used, i.e.

[0139] ;

[0140] Where a, b, and c are the three vertices of the triangle, and a feature has three membership functions to describe the low, medium, and high levels of that feature.

[0141] The first layer can be described as follows:

[0142] ;

[0143] The second layer uses membership degrees to calculate the excitation intensity of the samples based on all rules. In this paper, all sub-ANFIS models have four input features, and the model is rule-based. The output of the second layer is the product of the membership degrees of all features:

[0144] ;

[0145] The third layer completes the normalization of the output of the previous layer:

[0146] ;

[0147] It is the normalized excitation intensity.

[0148] The fourth layer calculates the results of all rules, using a linear function to combine the inputs:

[0149] ;

[0150] Finally, the fifth layer completes the deblurring process:

[0151] ;

[0152] For ANFIS, the result of a sample calculation is a floating-point number, but what is needed is a binary 0-1 variable, so rounding can solve this problem.

[0153] Ensemble learning

[0154] Ensemble learning combines multiple independent models to improve generalization performance. Three strategies in current ensemble learning techniques are bagging, boosting, and stacking.

[0155] Unlike most models, the ANFIS model yields twice the expected result, defining a result above 0.5 as positive and below 0.5 as negative. In this paper, several bagging strategies will be explored:

[0156] (1) Voting mode

[0157] ;

[0158] ;

[0159] Where m is the number of sub-models. In CIR-ANFIS, m is the number of features minus 3. res is the prediction of each sub-model, and the result is the prediction of the entire model. If more than or equal to 50% of the sub-models correctly predict a sample as a positive class, the final result is positive.

[0160] (2) Persuasion Model

[0161] ;

[0162] The variables have the same meaning as described above. When using this strategy, if a sub-model strongly favors the positive or negative class, a single positive or negative class may affect the result of the entire model.

[0163] (3) Persuasion pattern with inhibition function

[0164] ;

[0165] Use continuous functions to reduce the negative impact of extreme values ​​on classification performance, including square root functions and logarithmic functions:

[0166] ;

[0167] ;

[0168] (4) Semi-persuasion model

[0169] ;

[0170] Here, 'a' and 'b' are upper and lower bounds, which limit the ability to "persuade" other sub-models. They can be floating-point numbers or infinity. In this paper, two semi-persuasion methods will be attempted: 'a' is 0 and 'b' is infinity, or 'a' is 0 and 'b' is 1.

[0171] After five steps, the CIR-ANFIS model is obtained.

[0172] The features of this implementation plan are: (1) a CIR-NFIS model is proposed, which can be applied to the diagnosis and prediction of diabetes, and can also be used to classify binary disease prediction scenarios, especially prediction scenarios based on laboratory indicators.

[0173] (2) Introducing causal reasoning into feature selection and training the model by randomly generating a subset of data with causal coefficients as weights reduces the model training cost and achieves better model performance. Some attributes and results in disease prediction based on laboratory indicators are likely to have causal relationships. Causal inference can help the model run better while maintaining its interpretability.

[0174] (3) Unlike conventional ensemble methods, a “semi-persuasive” approach is used to combine the prediction results of all sub-models;

[0175] (4) Use experimental data to demonstrate the performance of the model.

[0176] In this implementation, a dataset is created using Python, causal inference is performed, and normalization is carried out from the first to the third part. For the remaining part, an ANFIS sub-model is trained using MATLAB, ensemble learning is performed, and the results are computed. All tests are conducted on a Huawei computer with an Intel Core i7-10510 processor running at 2.30 GHz and 8 GB of memory.

[0177] Dataset

[0178] In this paper, we will use nine UCI datasets to test the performance of CIR-ANFIS. The features and sample size of each dataset are shown in Table 1. Ten cross-validations were performed on all datasets.

[0179] Table 1. Dataset Overview

[0180] Serial Number name Number of features Sample size Number of sub-models CIR-ANFIS Number of sub-models: 2CIR-ANFIS 1 PIMA 8 687 5 10 2 FURIN 17 1836 14 28 3 Tabular 10 442 7 14 4 Sylhet 16 520 13 26 5 LMCH 11 1000 8 16 6 Acute Inflammations 8 120 5 10 7 Breast cancer data 15 286 12 24 8 Breast Cancer Coimbra 9 116 6 12 9 Wisconsin Breast Cancer Database 9 684 6 12

[0181] Table 2. Experimental Accuracy

[0182] Model Number 1 2 3 4 5 6 7 8 9 Direct use of causal reasoning 0.8321 0.7750 0.7231 0.8056 0.9645 0.8235 0.5870 0.6471 0.9730 Recursive feature elimination 0.8030 0.7961 0.7615 0.9306 0.9590 1.0000 0.7007 0.8529 0.9838 vote 0.9243 0.7892 0.2757 0.5622 0.9405 0.6432 0.4162 0.4865 0.9189 persuade 0.9493 0.8172 0.7846 0.7778 0.9617 1.0000 0.6087 0.8529 0.9838 Logarithmic persuasion 0.9565 0.8158 0.7962 0.8819 0.9645 1.0000 0.5978 0.6765 0.9730 Square root persuasion 0.7754 0.7231 0.7231 0.7292 0.9645 1.0000 0.5870 0.6176 0.9730 0-Infinitely persuading 0.9565 0.8115 0.7769 0.8194 0.9645 1.0000 0.5978 0.7353 0.9730 0-1 semi-persuasion 0.9565 0.8242 0.7846 0.8958 0.9645 1.0000 0.6087 0.8235 0.9784 2x model voting 0.6812 0.5063 0.6308 0.8681 0.6448 0.5882 0.5543 0.6176 0.5189 2x model persuasion 0.9493 0.8678 0.7769 0.8194 0.9617 1.0000 0.5435 0.6765 0.9838 2x model log persuasion 0.9565 0.8523 0.7615 0.8681 0.959 1.0000 0.5543 0.6176 0.9784 2x model square root persuasion 0.9493 0.8608 0.7769 0.8681 0.9645 1.0000 0.5543 0.7059 0.9784 2x model 0-infinitely persuading 0.9493 0.8594 0.7692 0.8681 0.9645 1.0000 0.5543 0.7059 0.9784 2x model 0-1 half persuasion 0.9420 0.8706 0.8000 0.8681 0.9645 1.0000 0.5543 0.7647 0.9838

[0183] Table 3 Model G-mean

[0184] Model Number 1 2 3 4 5 6 7 8 9 Direct use of causal reasoning 0.8126 0.7740 0.6779 0.7779 0.9632 0.8672 0.6843 0.7839 0.9737 Recursive feature elimination 0.7950 0.8085 0.6763 0.9187 0.9578 1.0000 0.8274 0.8909 0.9847 vote 0.7426 0.5810 0.7942 0.8889 0.7477 0.7670 0.7445 0.7859 0.7204 persuade 0.9388 0.8164 0.8405 0.7386 0.9632 1.0000 0.7405 0.8840 0.9845 Logarithmic persuasion 0.9478 0.8153 0.8315 0.8585 0.9632 1.0000 0.7486 0.7835 0.9745 Square root persuasion 0.9478 0.8176 0.8359 0.7969 0.9632 1.0000 0.7422 0.8292 0.9745 0-Infinitely persuading 0.9478 0.8118 0.8359 0.7969 0.9632 1.0000 0.7422 0.8131 0.9745 0-1 semi-persuasion 0.9478 0.8223 0.8405 0.8752 0.9632 1.0000 0.7405 0.8645 0.9794 2x model voting 0.7568 0.7101 0.7942 0.8689 0.7620 0.7670 0.7445 0.7859 0.7204 2x model persuasion 0.9388 0.8801 0.8514 0.7706 0.9632 1.0000 0.7339 0.7572 0.9847 2x model log persuasion 0.9478 0.8684 0.8462 0.8402 0.9574 1.0000 0.7445 0.7979 0.9798 2x model square root persuasion 0.9388 0.8718 0.8541 0.8387 0.9632 1.0000 0.7445 0.8292 0.9798 2x model 0-infinitely persuading 0.9388 0.8730 0.8501 0.8387 0.9632 1.0000 0.7445 0.757 0.9798 2x model 0-1 half persuasion 0.9297 0.8813 0.8663 0.8387 0.9632 1.0000 0.7445 0.7979 0.9847

[0185] Table 4 Model F1 Score

[0186] Model Number 1 2 3 4 5 6 7 8 9 Direct use of causal reasoning 0.8642 0.7740 0.7882 0.7778 0.9625 0.8636 0.6780 0.7692 0.9735 Recursive feature elimination 0.8428 0.8074 0.8088 0.9167 0.9577 1.0000 0.8257 0.8889 0.9846 vote 0.7108 0.5613 0.7736 0.8828 0.7171 0.7407 0.7133 0.7636 0.683 persuade 0.9369 0.8164 0.8391 0.7377 0.9595 1.0000 0.7237 0.8837 0.9845 Logarithmic persuasion 0.9464 0.8152 0.8295 0.8547 0.9625 1.0000 0.7259 0.7755 0.9744 Square root persuasion 0.7832 0.7587 0.8022 0.7310 0.9625 1.0000 0.7206 0.7547 0.9744 0-Infinitely persuading 0.9464 0.8118 0.8343 0.7969 0.9625 1.0000 0.7218 0.8085 0.9744 0-1 semi-persuasion 0.9565 0.8242 0.7846 0.8958 0.9645 1.0000 0.6087 0.8235 0.9784 2x model voting 0.7284 0.6704 0.7736 0.8652 0.7347 0.7407 0.7133 0.7636 0.6833 2x model persuasion 0.9493 0.8678 0.7769 0.8194 0.9617 1.0000 0.5435 0.6765 0.9838 2x model log persuasion 0.9565 0.8523 0.7615 0.8681 0.9590 1.0000 0.5543 0.6176 0.9784 2x model square root persuasion 0.9493 0.8608 0.7769 0.8681 0.9645 1.0000 0.5543 0.6176 0.9784 2x model 0-infinitely persuading 0.9493 0.8594 0.7692 0.8681 0.9645 1.0000 0.5543 0.7059 0.9784 2x model 0-1 half persuasion 0.9420 0.8706 0.8000 0.8681 0.9645 1.0000 0.5543 0.7647 0.9838

[0187] Table 5 Model Recall

[0188] Model Number 1 2 3 4 5 6 7 8 9 Direct use of causal reasoning 0.8974 0.7806 0.8171 0.7903 1.0000 0.7917 0.5970 0.6452 0.9892 Recursive feature elimination 0.8590 0.7677 0.9146 0.9821 0.9714 1.0000 0.7759 0.8333 0.9697 vote 0.5514 0.4463 0.6308 0.7901 0.5590 0.5882 0.5543 0.6176 0.5189 persuade 1.0000 0.8234 0.7935 0.7759 1.0000 1.0000 0.5949 0.8261 0.9794 Logarithmic persuasion 1.0000 0.8210 0.7766 0.9434 1.0000 1.0000 0.5833 0.6786 0.9596 Square root persuasion 1.0000 0.8257 0.7849 0.7969 1.0000 1.0000 0.5854 0.7600 0.9596 0-Infinitely persuading 1.0000 0.8141 0.7849 0.7949 1.0000 1.0000 0.5854 0.7308 0.9596 0-1 semi-persuasion 1.0000 0.8353 0.7935 0.9804 1.0000 1.0000 0.5949 0.8636 0.9694 2x model voting 0.5728 0.5042 0.6308 0.7922 0.5806 0.5882 0.5543 0.6176 0.5189 2x model persuasion 1.0000 0.8668 0.7524 1.0000 1.0000 1.0000 0.5495 0.7083 0.9697 2x model log persuasion 1.0000 0.8439 0.7339 0.9412 1.0000 1.0000 0.5543 0.6333 0.9600 2x model square root persuasion 1.0000 0.9507 0.7477 0.9787 1.0000 1.0000 0.5543 0.7037 0.9600 2x model 0-infinitely persuading 1.0000 0.8529 0.7407 0.9787 1.0000 1.0000 0.5543 0.7037 0.9600 2x model 0-1 half persuasion 1.0000 0.8692 0.7692 0.9787 1.0000 1.0000 0.5543 0.7600 0.9697

[0189] Table 6 Model Accuracy

[0190] Model Number 1 2 3 4 5 6 7 8 9 Direct use of causal reasoning 0.8333 0.7675 0.7576 0.7656 0.9278 0.9499 0.7844 0.9525 0.9582 Recursive feature elimination 0.8272 0.8515 0.7614 0.8594 0.9444 1.0000 0.8823 0.9522 1.0000 vote 1 0.7563 1 1 1 1 1 1 1 persuade 0.8814 0.8095 0.8902 0.7031 0.9222 1.0000 0.9216 0.9048 0.9896 Logarithmic persuasion 0.8983 0.8095 0.8902 0.7969 0.9278 1.0000 0.9608 0.9048 0.9896 Square root persuasion 0.8983 0.8095 0.8902 0.7813 0.9278 1.0000 0.9412 0.9048 0.9896 0-Infinitely persuading 0.8983 0.8095 0.8902 0.7813 0.9278 1.0000 0.9412 0.9048 0.9896 0-1 semi-persuasion 0.8983 0.8095 0.8902 0.7969 0.9278 1.0000 0.9216 0.9048 0.9896 2x model voting 1 1 1 0.9531 1 1 1 1 1 2x model persuasion 0.8814 0.8936 0.9634 0.5938 0.9222 1.0000 0.9804 0.8095 1.0000 2x model log persuasion 0.8983 0.8936 0.9756 0.7500 0.9167 1.0000 1.0000 0.9048 1.0000 2x model square root persuasion 0.8814 0.8936 0.9756 0.7188 0.9278 1.0000 1.0000 0.9048 1.0000 2x model 0-infinitely persuading 0.8814 0.8936 0.9756 0.7188 0.9278 1.0000 1.0000 0.9048 1.0000 2x model 0-1 half persuasion 0.8644 0.8936 0.9756 0.7188 0.9278 1.0000 1.0000 0.9048 1.0000

[0191] Evaluation indicators and parameter settings

[0192] To comprehensively evaluate the performance of CIR-ANFIS, five metrics are used to describe its performance: accuracy, F1 score, G-means, precision, and recall. Considering binary classification, based on the confusion matrix, the five metrics can be calculated as follows:

[0193] Accuracy determines the percentage of samples that are correctly predicted out of all samples.

[0194] ;

[0195] Precision can find the ratio of true class instances in all positive class matches.

[0196] ;

[0197] Recall is defined as the number of correctly predicted positive class samples divided by the actual number of positive class samples.

[0198] ;

[0199] The G-mean and F1 score are defined as follows:

[0200] ;

[0201] ;

[0202] We applied the five ensemble learning methods listed in the CIR-ANFIS methodology section and compared them to using causal inference only on all datasets. We did not create 2CIR-ANFIS for the remaining two datasets because there were insufficient features to create two separate subsets.

[0203] Experimental results:

[0204] Tables 2 to 6 show the accuracy, precision, recall, G-mean, and F1 score of the model across all nine datasets.

[0205] The table shows that CIR-ANFIS using semi-persuasion achieves the best accuracy on 7 out of 11 datasets when a is 0 and b is 1, particularly on the tabular dataset, where it improves classification accuracy by more than 10 times. For the remaining 4 datasets, the difference between the best score and 0-1 semi-persuasion is negligible. Overall, utilizing CIR-ANFIS can improve efficiency while maintaining interpretability when creating ANFIS models through causal inference.

[0206] The number of sub-models in 2CIR-ANFIS and CIR-ANFIS was compared. According to the experimental results, 2CIR-ANFIS performs slightly better than CIR-ANFIS on datasets with more than 10 features, while CIR-ANFIS performs better on datasets with fewer than 10 features.

[0207] The above technical solutions only embody the preferred technical solutions of the present invention. Any modifications that may be made by those skilled in the art to certain parts thereof embody the principles of the present invention and fall within the protection scope of the present invention.

Claims

1. An adaptive neurofuzzy system for diabetes analysis and integration, characterized in that, This includes step S1: inputting the dataset into the CIR-ANFIS model, where the model first standardizes the dataset; Step S2: Calculate the causal coefficient of each feature corresponding to the prediction result using causal reasoning, and use this coefficient as the feature weight; Step S3: Randomly select multiple different subsets, construct an adaptive fuzzy inference system for each different subset, and obtain the prediction results of each inference system; Step S4: Obtain the prediction results of the entire model through ensemble learning; Causal effect can be considered as the influence of each feature on classification. Features with high levels of causal influence may contribute more than features with low levels of causal influence. Using causal effect as feature weights, random subsets are generated through the following operations; In the CIR-ANFIS model, using There are 4 sub-models, where n is the number of input features. Each sub-model has four input features. For the 2CIR-ANFIS model, twice the number of input features is used. The following is an algorithm for selecting the input features for each sub-model: (1) For each feature, the weight score The definition is as follows: ； in It is the absolute value of the causal relationship between features and results. It is a random number between 0 and 1. (2) Select the four features with the highest weight scores. (3) If this is not the first sub-model, compare the selection result with the previous result. If all four features match, the selection result is invalid; return to step 1 and perform the calculation again; otherwise, continue to calculate each subsequent sub-model until all sub-data has been calculated. Through the above process, the steps of preparing the sub-dataset for sub-model training are completed. To demonstrate that this method can use more features, it is compared with the traditional simple selection of four features. In CIR-ANFIS, for those with The probability of a dataset with 1 feature not being selected by the submodel. It can be approximated as: ； The probability of not being selected by all sub-models for: ； Continue the analysis under various conditions; (1) It is significantly higher than most other common traits, ranking in the top 4. This characteristic is likely to be included in the subset multiple times, which means that it can contribute to many sub-models. Obviously, it makes a great contribution to the entire CIR-ANFIS model. (2) If the ranking is below or above the Top-4, but the difference is not significant, then... The value ranges from 0.1 to 0.

2. The value ranges from 0.4 to 0.6, which indicates that this feature can still be incorporated into some subsets of the dataset. In other words, the CIR-ANFIS model allows for the inclusion of more features rather than just selecting a limited number of features. (3) Significantly lower than the Top-4, being one-twentieth or even one-hundredth of the Top-4, in this case... Below 0.05, A value higher than 0.8 means that due to the low causal relationship between features and results, features may sometimes only be included in one or a few subsets, or even not in any subsets at all. Based on the above calculations and analysis, CIR-ANFIS can use causal inference to create subset datasets that reflect the relative importance of various features, with the aim of using more data from more features and improving the performance of the overall model.

2. The adaptive neurofuzzy system for diabetes analysis and integration according to claim 1, characterized in that, In step S1, the input dataset needs to be standardized before training the CIR-ANFIS model. In this section, all features are converted to a form with a mean of 0 and a variance of 1. The steps are as follows: (1) For each feature, calculate its mean and then subtract its mean; (2) Calculate the variance of each feature, and then divide the result of (1) by the variance.

3. The adaptive neurofuzzy system for diabetes analysis and integration according to claim 1, characterized in that, The weight of each feature in step S2 may be determined by its causal relationship with the outcome; this indication is called causal effect. In statistics, causal inference is the experimental study that determines causal relationships. The following are the four steps of causal inference: Step 1: Model the problem based on assumptions; Since causal models are developed using problem assumptions, the prior knowledge used in their construction can be used to incorporate past information and create a graph. Without prior knowledge of the dataset, it is assumed that there is no causal relationship between features and all features are affected by unobserved confounding factors. When determining the causal relationship between a specific feature and the outcome, it is assumed that if the feature is above the mean and the outcome is positive, then there is a causal relationship. To this end, normalized features are converted into 0-1 variables and the causal relationship is calculated using these variables. If there is prior knowledge of the dataset, this knowledge can be used to create a more refined causal inference graph, which helps to obtain more accurate results. Step 2: Determine the causal relationship; The definition of causal analysis indicates that changes in an intervention will affect the outcome when other variables remain constant. After determining the causal effect of the causal model, it can be estimated based on previous assumptions. In this study, the backdoor criterion was mainly used. Step 3: Use statistical methods to estimate; In this step, the causal relationship between each feature and outcome can be calculated. Many techniques, such as propensity score matching, propensity score stratification, propensity score-based inverse probability weighting, linear regression, generalized linear models, instrumental variables, and point regression, have been used to determine statistical causal effects. The causal effect ranges between -1 and 1, where a negative value indicates a negative correlation between the feature and the outcome. Step 4: Use multiple robustness testing methods to verify the credibility of the causal effect; Although statistical estimations of the data were used in Section 3 to calculate causality, the causality itself is based on prior assumptions rather than the data. Therefore, several robustness tests were employed to demonstrate the prior assumptions. The three robustness tests used in this study are as follows: 1) Random confounding factor; In this test, independent random variables are introduced as common influencing factors of the dataset. If the assumptions are accurate, the estimation results will not change naturally. 2) Data subset validation; In this test, a randomly selected subset is used instead of the provided dataset, and if the assumptions are true, the estimation results will not change significantly. 3) Placebo intervention; In this test, independent random variables were used instead of the actual intervention variables; If the assumptions hold true, the estimate will be close to zero.

4. The adaptive neurofuzzy system for diabetes analysis and integration according to claim 1, characterized in that, The ANFIS mentioned is a classic TSK fuzzy model, which adopts a Takagi-Sugeno-Kang type first-order rule, defined as follows: ； in, It is the i-th feature of the data. It is a fuzzy set, and y is the result of the rule. These are the result coefficients of the TSK rule. The result part in ANFIS is a linear function of the input vector, rather than a fuzzy rule of the Mamdani type. Using a linear function makes ANFIS easier to understand compared to other more complex functions. A typical ANFIS consists of 5 layers; In the first layer, membership functions are used to blur the input samples. The number of membership functions can be adjusted before training the ANFIS model. Various function shapes are widely used in this layer; in this study, triangular membership functions are used, i.e. ； Where a, b, and c are the three vertices of the triangle, and a feature has three membership functions to describe the low, medium, and high levels of the feature; The first layer can be described as follows: ； The second layer uses membership degrees to calculate the excitation intensity of samples based on all rules. In this paper, all sub-ANFIS models have four input features, and the model is rule-based. The output of the second layer is the product of the membership degrees of all features. ； The third layer completes the normalization of the output of the previous layer: ； It is the normalized excitation intensity; The fourth layer calculates the results of all rules, using a linear function to combine the inputs: ； Finally, the fifth layer completes the deblurring process: ； For ANFIS, the result of a sample calculation is a floating-point number, but what is needed is a binary 0-1 variable, so rounding can solve this problem.

5. The adaptive neurofuzzy system for diabetes analysis and integration according to claim 1, characterized in that, To improve generalization performance, ensemble learning combines multiple independent models. The three strategies of ensemble learning techniques today are bagging, boosting, and stacking. Unlike most models, the ANFIS model yields twice the expected result, defining a result above 0.5 as positive and below 0.5 as negative. Several bagging strategies will be tested: (1) Voting mode ； ； Where m is the number of sub-models, and in CIR-ANFIS, m is the number of features minus 3; res is the prediction of each sub-model. The result is the prediction of the entire model. If more than or equal to 50% of the sub-models correctly predict the positive class result of a sample, the final result is positive. (2) Persuasion Model ； The variables have the same meaning as above. When using this strategy, if a sub-model strongly tends to be positive or negative, a positive or negative class may affect the result of the entire model. (3) Persuasion pattern with inhibition function ； Use continuous functions to reduce the negative impact of extreme values ​​on classification performance, including square root functions and logarithmic functions: ； ； (4) Semi-persuasion model ； Where a and b are upper and lower bounds, which limit the ability to "persuade" other sub-models. They are floating-point numbers or infinity. In this paper, we will try two semi-persuasion methods: a is 0 and b is infinity, or a is 0 and b is 1. After five steps, the CIR-ANFIS model is obtained.

Images