A hierarchical fuzzy neural network incremental modeling method and device for classifying metabolic fatty liver data

Abstract

This invention proposes a hierarchical fuzzy neural network incremental modeling method and apparatus for classifying metabolic fatty liver data, belonging to the field of data processing technology. First, biochemical index data associated with metabolic fatty liver are collected as features to form a metabolic fatty liver dataset. A feature importance evaluation index is established, and the features are sorted in descending order and divided into feature data subsets. Then, the number of hierarchical structures in the hierarchical model is determined, and the feature data subsets input to each substructure are identified. A cascaded hierarchical model is constructed using fuzzy neural networks and incremental learning methods, forming a multi-stage hierarchical incremental parameter learning strategy. Finally, the parameter learning process of the subsequent part of the hierarchical fuzzy neural network is constructed to obtain a transferable hierarchical fuzzy neural network model, yielding prediction results for metabolic fatty liver data. This invention possesses excellent generalization ability, stability, and modeling speed, making it a feasible intelligent algorithm model for processing metabolic fatty liver disease data.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to the field of artificial intelligence (artificial neural network) pattern data fitting, recognition, and classification. More specifically, it relates to an incremental modeling method using a hierarchical fuzzy neural network (HFNN) for the classification of metabolic fatty liver data. Background Technology

[0002] The analysis and classification of metabolic fatty liver feature data based on artificial intelligence methods can provide relatively sufficient correlation and reference information on disease characteristics, which has important auxiliary application significance and clinical value for the diagnosis and subsequent treatment of metabolic fatty liver. Metabolic fatty liver feature data is characterized by complex distribution, diverse correlation information, and diverse data sources, placing certain requirements on the artificial intelligence model used for classification tasks. This invention proposes an incremental hierarchical fuzzy neural network learning algorithm to achieve feature analysis and classification processing of metabolic fatty liver data, thereby achieving the purpose of assisting in the diagnosis of metabolic fatty liver disease.

[0003] Incremental learning is a technique that continuously learns and adapts to new knowledge based on existing knowledge, allowing intelligent models to integrate new data without retraining. Compared to traditional batch learning, incremental learning is more flexible, can dynamically adjust in changing environments, avoids forgetting previous knowledge, and continuously improves performance by receiving new data. Therefore, it has been widely used in machine learning and artificial intelligence.

[0004] Hierarchical network models are a feasible method for processing complex, high-dimensional, and large-scale data. By analyzing the correlation and referential information of data features, initial feature segmentation is completed, followed by multi-stage learning through a serial hierarchical model structure. This serial hierarchical model structure consists of multiple hierarchical fuzzy neural network models, which progressively input segmented feature data layer by layer, while maintaining connections with the data and learning results of the previous layer to jointly complete the multi-stage parameter learning task.

[0005] With the deepening of research on metabolic fatty liver disease and related topics, in addition to the main biological indicators of metabolic fatty liver disease, diseases such as diabetes, hypertension, and hyperuricemia, as well as lifestyle factors such as obesity, alcohol consumption, and smoking, have also been included in the scope of etiological investigation of metabolic fatty liver disease. This has resulted in the characteristic data used to describe metabolic fatty liver disease being diverse in distribution, rich in correlation information, and from a wide range of data sources, forming a complex, high-dimensional, and large-scale metabolic fatty liver disease dataset. How to conduct data analysis and perform classification tasks on the diversified characteristic data of metabolic fatty liver disease, and combine artificial intelligence methods to achieve auxiliary diagnosis and risk assessment of metabolic fatty liver disease, is a current hot research topic and an urgent problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a hierarchical fuzzy neural network incremental modeling method for classifying characteristic data of metabolic fatty liver disease. While ensuring the generalization ability and modeling performance of the fuzzy neural network model, the method takes into account the characteristics of the characteristic data of metabolic fatty liver disease through a multi-stage hierarchical learning strategy established by the hierarchical structure, and improves the learning ability of the model through the incremental learning strategy, so as to realize the auxiliary diagnosis and risk assessment of metabolic fatty liver disease.

[0007] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0008] A hierarchical fuzzy neural network incremental modeling method for classifying metabolic fatty liver data includes the following steps:

[0009] S1. Collect data and normalize the data to form a data sample set of metabolic fatty liver;

[0010] S2. Based on the metabolic fatty liver data sample set, feature importance analysis is performed using importance evaluation indicators. Based on the analysis results, the data samples are divided into multiple feature data subsets.

[0011] S3. Based on the feature data subset partitioning results, adaptively determine the number of cascaded hierarchical model structures of the incremental hierarchical fuzzy neural network, and select feature data subsets input into each level of the model;

[0012] S4. Based on the fuzzy neural network modeling strategy and incremental learning mechanism, construct the base models at each level in the hierarchical model structure to form a cascaded hierarchical incremental learning model.

[0013] S5. Based on the cascaded hierarchical incremental learning model, perform parameter learning to obtain a transferable incremental hierarchical fuzzy neural network model, and then carry out data classification of metabolic fatty liver disease.

[0014] In addition to the aspects and any possible implementations described above, an implementation is further provided, wherein S1 specifically includes:

[0015] A data sample set is formed by collecting biochemical indicator test data related to metabolic fatty liver disease. The input sequence of the data sample set is represented as follows: It contains characteristic information describing metabolic fatty liver disease, and the output sequence is represented as follows: It contains label information describing the diagnosis of metabolic fatty liver disease; N is the number of samples, d is the number of features, and k is the number of labels.

[0016] Map each data point in the data sample set to [0, 1] to form a normalized data sample. The normalization formula is as follows:

[0017]

[0018] In the formula, X ij Let X be the original data point in the i-th row and j-th column of the data sample set, min(X) j ) is the minimum value of the feature belonging to the j-th column in the data sample set, max(X) j Let be the maximum value of the feature in the j-th column of the data sample set, where i = 1, 2, ..., N, j = 1, 2, ..., d. The above formula ensures that the value of each feature is within the range [0, 1], while maintaining the feature distribution of each group of samples.

[0019] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S2 specifically includes:

[0020] Based on the metabolic fatty liver data sample set described in S1, and combined with the following probability model, an importance evaluation index is established:

[0021]

[0022] In the formula, P(X) j ) represents the importance evaluation index of the feature data in the j-th column of the data sample. For the data sample set, X j It is the feature information of the j-th column in the data sample set. Let j = 1, 2, ..., d, where ||.||2 represents the Euclidean distance. The p-value represents the importance of each feature in the data sample. It quantifies the correlation between each feature and the label information. As the p-value increases, the strength of the correlation or influencing factor gradually increases, that is, the increase in p-value is positively correlated with its importance.

[0023] Based on the aforementioned importance evaluation indicators, the importance of each column of feature data in the data sample is calculated, and the data is sorted in descending order according to the importance evaluation indicators. This completes the data feature importance analysis results, dividing the data sample into... The feature data subset is the feature columns corresponding to the top 30% of the data. The feature columns corresponding to the next 30% are used as a subset of feature data. The feature columns corresponding to the last 40% are used as a subset of feature data. N is the number of samples, d is the total number of features, and d1, d2, and d3 represent the number of features contained in each subset, where d1 + d2 + d3 = d. Based on the established importance evaluation index, the importance of the features represented by the feature data subsets X1, X2, and X3 decreases sequentially.

[0024] In addition to the aspects and any possible implementations described above, a further implementation is provided, wherein S3 specifically includes:

[0025] Based on the partitioning results of the metabolic fatty liver disease dataset, the original data sample set was divided into three feature data subsets X1, X2, and X3 after feature sorting. The hierarchical network with a three-layer cascade structure was represented in the following form:

[0026] y1=g1(X1), y2=g2(y1,X2), y3=g3(y2,X3),

[0027] That is:

[0028] y3 = g3[g2(g1(X1), X2), X3]

[0029] In the formula, y1 represents the output of the first layer in the three-layer cascaded structure, and its input is the feature data subset X1. g1(X1) represents the first layer substructure. y2 represents the output of the second layer in the three-layer cascaded structure, and its input is the feature data subset X2 and the output of the first layer substructure. g2(y1, X2) represents the second layer substructure. y3 represents the output of the third layer in the three-layer cascaded structure, which is also the output of the last layer. Its input is the feature data subset X3 and the output of the second layer substructure. g3[g2(g1(X1), X2), X3] represents the third layer substructure.

[0030] In addition to the aspects and any possible implementations described above, an implementation is further provided, wherein S4 specifically includes:

[0031] Based on the hierarchical network structure described above, and combining incremental learning strategies and fuzzy neural networks, a hierarchical fuzzy neural network incremental modeling strategy is constructed. First, a fuzzy neural network is constructed to establish the substructure of the hierarchical network model. The m-th fuzzy rule in this fuzzy neural network is represented as follows:

[0032] Rule m: IF X1 is A 1m …and X i is A im THAN T m isβ m m = 1, 2, ..., K

[0033] Among them, X i Let A be the i-th dimension of the input variable. im T is the membership function of the m-th rule with respect to the i-th input variable. m It is the output of the m-th rule, β m Let A be the consequent parameter of the m-th rule, and K be the number of fuzzy rules. The constructed membership function A... im As shown below:

[0034]

[0035] Among them, c im and σ im These represent the center and width of the membership function, respectively.

[0036] The output of fuzzy regularization is represented as:

[0037]

[0038] The above formula shows that the fuzzy rule layer contains K fuzzy rules.

[0039] Secondly, a cascaded hierarchical incremental learning model with a three-layer cascaded structure is established. In the three-layer cascaded structure, feature data subsets divided based on importance evaluation indicators are fed into each level of the model in batches.

[0040] In a three-layer cascaded structure, the learning process of the first-layer substructure can be represented as follows:

[0041]

[0042] In the formula, g1 is the Sigmoid activation function, and h1 is the fuzzy rule construction process, containing K1 fuzzy rules. and The input weights are randomly generated. and The input bias is randomly generated, and L1 is the number of neurons used. The learning result of the feature mapping node part is... The result of incremental learning is Together, they constitute the output of the first-layer substructure. y1 is and The result of the second-dimensional concatenation contains a total of 2L1 neuron nodes. Furthermore, and The values ​​are generated from random numbers uniformly distributed on [-λ1, λ1], where λ1 = [0.5, 250]. To ensure the convergence performance of the model, the configuration process of w1 and b1 is constrained by the following hierarchical supervision mechanism:

[0043]

[0044] in, yes The elements in the l1-th column, l1 = 1, 2, ..., L1, as well as Select The corresponding w1 and b1 serve as the input bias and weight of this substructure. When l1 = L1, the configuration process of w1 and b1 is completed.

[0045] The learning process of the second-layer substructure is represented as follows:

[0046]

[0047]

[0048] In the formula, g2 is the Sigmoid activation function, and h2 is the fuzzy rule construction process, containing K2 fuzzy rules. and The input weights are randomly generated. and The input bias is randomly generated, and L2 is the number of neurons used. The learning result of the feature mapping node part is... and The result of incremental learning is The output of the second layer was determined together. It contains a total of 3L2 neuron nodes. Furthermore, and Generated from random numbers uniformly distributed on [-λ2, λ2], where λ2 = [0.5, 250]. The configuration process of w2 and b2 is constrained based on the following hierarchical supervision mechanism:

[0049]

[0050] in, yes The element in column l2, where l2 = 1, 2, ..., L2, as well as Select The corresponding w2 and b2 serve as the input bias and weight of this substructure. When l2 = L2, the configuration process of w2 and b2 is completed.

[0051] The learning process of the third-layer substructure can be represented as:

[0052]

[0053] In the formula, g3 is the Sigmoid activation function, and h3 is the fuzzy rule construction process, containing K3 rules. and The input weights are randomly generated. and L3 is the randomly generated input bias, and L3 is the number of neurons used. The learning result of the feature mapping node part is... and The result of incremental learning is The output of the third layer was determined together. It contains a total of 3L3 neurons. y3, as the output of the hidden layer of the incremental hierarchical fuzzy neural network, is re-represented as... Its dimensions are N×L, where L = 2L1 + 3(L2 + L3). Furthermore, and The values ​​are generated from random numbers uniformly distributed on [-λ3, λ3], where λ3 = [0.5, 250]. The configuration process of w3 and b3 is constrained through the following hierarchical supervision strategy to ensure the convergence performance of the model:

[0054]

[0055] in, yes The element in column l3, where l3 = 1, 2, ..., L3, as well as Select The corresponding w3 and b3 serve as the input bias and weight of this substructure. When l3 = L3, the configuration process of w3 and b3 is completed.

[0056] In addition to the aspects and any possible implementations described above, an implementation is further provided, wherein S5 specifically includes:

[0057] The output of the hidden layer is obtained based on the hidden layer construction method and feature learning process of the incremental hierarchical fuzzy neural network described in S4. Solve for the output and output weights of the incremental hierarchical fuzzy neural network, and design the following objective loss function:

[0058]

[0059] In the formula, H is the output of the hidden layer of the incremental hierarchical fuzzy neural network, β is the output weight, T is the target object, γ1 is the Lasso regularization tuning parameter, γ2 is the Ridge regularization tuning parameter, ||.||1 represents the Manhattan distance, and ||.||2 represents the Euclidean distance. The Elastic-Net regularization strategy is composed of two different regularization methods.

[0060] The output weight β of the incremental hierarchical fuzzy neural network can be expressed as follows:

[0061] β=(H T H+γ2I)+ (H T T+γ1I)

[0062] In particular,

[0063]

[0064] In the formula, H + This represents the pseudo-inverse operation, where I is the identity matrix.

[0065] Based on the above, if metabolic fatty liver disease is used as the training dataset, then the output of the incremental hierarchical fuzzy neural network will... It includes predictions for metabolic fatty liver. It can be represented as follows:

[0066]

[0067] In the formula, H is the hidden layer output matrix of the incremental hierarchical fuzzy neural network, and β is the output weight.

[0068] The designed incremental hierarchical fuzzy neural network is primarily used for metabolic fatty liver disease data samples to detect metabolic fatty liver disease by performing a classification task. Therefore, the root mean square error (RMSE), accuracy (ACC), and area under the ROC curve (AUC) are used as performance evaluation metrics, which are expressed as follows:

[0069]

[0070]

[0071]

[0072] In the formula, AUC is the area under the ROC curve (Receiver Operating Characteristic Curve). The ROC curve is a curve with the False Positive Rate (FPR) on the horizontal axis and the True Positive Rate (TPR) on the vertical axis. Each point on the ROC curve represents the value of the True Positive Rate and the False Positive Rate at a different classification threshold. The value of AUC can be calculated by integrating the area under the ROC curve. This value is a commonly used indicator for evaluating the performance of classifiers.

[0073] This invention also provides an apparatus for analyzing and modeling characteristic data of metabolic fatty liver disease, comprising:

[0074] The data acquisition unit is used to collect data and normalize the data to form a sample of metabolic fatty liver characteristic data.

[0075] The first construction unit is used to perform feature analysis based on the metabolic fatty liver disease data samples combined with a probability model, and to divide the data samples into multiple feature data subsets according to the correlation analysis structure between data features.

[0076] The second construction unit is used to determine the cascaded hierarchical model structure of the incremental hierarchical fuzzy neural network and the input samples at each level based on the feature data subset partitioning results;

[0077] The third building unit is used to construct hierarchical model substructures, forming the hidden layers of the incremental hierarchical fuzzy neural network;

[0078] The computational unit learns parameters based on the cascaded hierarchical fuzzy neural network model and incremental learning strategy, and finally obtains the input layer weights, output weights and output sequence of the incremental hierarchical fuzzy neural network, and uses this to obtain the predicted output sequence, thereby obtaining the prediction result of metabolic fatty liver.

[0079] The beneficial effects of this invention are as follows: This invention proposes a novel incremental hierarchical fuzzy neural network modeling method, which can achieve the task of fitting and classifying input feature data through importance analysis and feature learning of metabolic fatty liver disease feature data. Metabolic fatty liver disease feature data is characterized by complex distribution, diverse associated data, and multiple data channels, forming a complex, high-dimensional, and large-scale sample set of lipid metabolism fatty liver disease. The proposed incremental hierarchical fuzzy neural network can perform importance analysis on each group of features through importance evaluation indicators, obtaining the degree of correlation and importance between each group of features and the target output. The feature set is sorted in descending order, thereby dividing the data samples as input data for the cascaded structure. The substructure of the incremental hierarchical fuzzy neural network consists of a fuzzy neural network model, which has the advantages of strong approximation ability, high model accuracy, and good generalization performance. Combining the advantages of hierarchical model structure in processing high-dimensional complex data, and the incremental learning strategy, the incremental hierarchical fuzzy neural network inherits the advantages of hierarchical models and fuzzy neural network models while having faster modeling speed, lower computational cost, and stronger learning ability, making it suitable for modeling tasks on metabolic fatty liver disease feature datasets. With the same model parameters, the modeling speed of this method is positively correlated with the number of hierarchical structures. Furthermore, because this method includes an importance analysis process based on importance evaluation indicators, combined with the modeling capabilities of fuzzy neural networks, it can be applied to different types of metabolic fatty liver disease (MAD) characteristic data and applications. It can also add more diverse biochemical indicators and lesion features to expand the MAD feature set, thereby improving the detection and risk assessment of MAD. This further ensures the scalability of using incremental hierarchical fuzzy neural networks for data analysis and classification tasks on MAD datasets, simplifies the modeling difficulty of MAD data, and provides a feasible solution for the auxiliary diagnosis and risk assessment of MAD. Attached Figure Description

[0080] Figure 1 This is a schematic diagram of an incremental hierarchical fuzzy neural network for detecting metabolic fatty liver disease provided by the present invention;

[0081] Figure 2 This is a schematic diagram of the hierarchical structure of a multi-stage incremental learning model provided by the present invention;

[0082] Figure 3 This invention provides a process for detecting and assessing the risk of metabolic fatty liver disease based on an incremental hierarchical fuzzy neural network for a specific dataset.

[0083] Figure 4 This is a schematic block diagram of a metabolic fatty liver data classification feature device provided by the present invention. Detailed Implementation

[0084] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0085] In view of this, embodiments of the present invention provide a hierarchical fuzzy neural network incremental modeling method for classifying metabolic fatty liver data. Please refer to... Figure 1 The flowchart illustrating a hierarchical fuzzy neural network incremental modeling method provided in this embodiment of the invention includes:

[0086] S1. Collect data and normalize the data to form a data sample set of metabolic fatty liver;

[0087] S2. Based on the metabolic fatty liver data sample set, feature importance analysis is performed using importance evaluation indicators. Based on the analysis results, the data samples are divided into multiple feature data subsets.

[0088] S3. Based on the feature data subset partitioning results, adaptively determine the number of cascaded hierarchical model structures of the incremental hierarchical fuzzy neural network, and select feature data subsets input into each level of the model;

[0089] S4. Based on the fuzzy neural network modeling strategy and incremental learning mechanism, construct the base models at each level in the hierarchical model structure to form a cascaded hierarchical incremental learning model.

[0090] S5. Based on the cascaded hierarchical incremental learning model, perform parameter learning to obtain a transferable incremental hierarchical fuzzy neural network model, and then carry out data classification of metabolic fatty liver disease.

[0091] Specifically, S1 includes:

[0092] Based on biochemical indicators associated with metabolic fatty liver disease and other numerical indicators of patients, the primary indicators were selected as the main characteristics describing metabolic fatty liver disease, namely: gender, age at physical examination, alanine aminotransferase (ALT), aspartate aminotransferase (AST), gamma-glutamyl transferase (GGT), total bilirubin (TBIL); uric acid (UA), total cholesterol (TC), triglycerides (TG), waist circumference (WC), body mass index (BMI), presence of diabetes, and high-density lipoprotein cholesterol (HDL-C). Data samples were collected from the above biochemical indicators associated with metabolic fatty liver disease, and the input sequence of the data sample set is represented as follows: It contains characteristic information describing metabolic fatty liver disease; the output sequence is represented as follows: It includes label information describing the diagnosis of metabolic fatty liver disease; N is the number of samples, d is the number of features, and k is the number of labels.

[0093] Map each data point in the data sample set to [0, 1] to form a normalized data sample. The normalization formula is as follows:

[0094]

[0095] In the formula, X ij Let X be the original data point in the i-th row and j-th column of the data sample set, min(X) j ) is the minimum value of the feature belonging to the j-th column in the data sample set, max(X) j Let d be the maximum value of the feature in the j-th column of the data sample set, where i = 1, 2, ..., N, j = 1, 2, ..., d. The above formula ensures that the value of each feature is within the range [0, 1], while maintaining the feature distribution of each group of samples. This yields the normalized feature dataset of metabolic fatty liver disease, completing the data preprocessing process.

[0096] Then, feature importance analysis was performed on the collected metabolic fatty liver dataset samples using S2, specifically including:

[0097] Based on the metabolic fatty liver dataset described in S1, and combined with the following probability model, an importance evaluation index is established:

[0098]

[0099] In the formula, P(X) j ) represents the importance evaluation index of the feature data in the j-th column of the data sample. For the data sample set, X j It is the feature information of the j-th column in the data sample set. Let j = 1, 2, ..., d, where ||.||2 represents the Euclidean distance. The p-value represents the importance of each feature in the data sample, quantifying the correlation between each feature and the label information. As the p-value increases, the strength of the correlation or influencing factor gradually increases; that is, the increase in p-value is positively correlated with its importance.

[0100] Based on the aforementioned importance evaluation indicators, the importance of each column of feature data in the data sample is calculated, and the data is sorted in descending order according to the importance evaluation indicators. This completes the data feature importance analysis and divides the data sample into [various categories]. The feature columns corresponding to the top 30% are used as feature data sub-columns. The feature columns corresponding to the next 30% are used as a subset of feature data. The feature columns corresponding to the last 40% are used as a subset of feature data. N is the number of samples, d is the total number of features, and d1, d2, and d3 represent the number of features contained in each subset, where d1 + d2 + d3 = d. Based on the established importance evaluation index, the importance of the features represented by the feature data subsets X1, X2, and X3 decreases sequentially.

[0101] Furthermore, based on the feature importance ranking and feature data subset partitioning derived in S2, an incremental hierarchical fuzzy neural network hierarchical structure is constructed through S3, such as... Figure 2 As shown, it specifically includes:

[0102] Based on the feature dataset partitioning results, the original data samples are divided into three feature data subsets X1, X2, and X3 after feature sorting. The hierarchical network with a three-layer cascade structure can be represented as follows:

[0103] y1=g1(X1), y2=g2(y1,X2), y3=g3(y2,X3)

[0104] That is:

[0105] y3 = g3[g2(g1(X1), X2), X3]

[0106] In the formula, y1 represents the output of the first layer in the three-layer cascaded structure, and its input is the feature data subset X1; g1(X1) represents the first-layer substructure. y2 represents the output of the second layer in the three-layer cascaded structure, and its input is the feature data subset X2 and the output of the first-layer substructure; g2(g1(X1), X2) represents the second-layer substructure. y3 represents the output of the third layer in the three-layer cascaded structure, which is also the output of the last layer; its input is the feature data subset X3 and the output of the second-layer substructure; g3[g2(g1(X1), X2), X3] represents the third-layer substructure. Thus, each sub-feature outputs a hierarchical structure in stages, while maintaining the connection between each level, thereby establishing a multi-stage hierarchical learning structure.

[0107] Furthermore, based on the multi-stage hierarchical learning structure established by S4, the learning process for incremental hierarchical fuzzy neural network parameters is constructed in S4, as specifically described below:

[0108] Based on the hierarchical network structure described in S3, and combining the incremental learning strategy and the fuzzy neural network, a hierarchical fuzzy neural network incremental modeling strategy is constructed. First, a fuzzy neural network is constructed to establish the substructure of the hierarchical network model. The j-th fuzzy rule in this fuzzy neural network is represented as follows:

[0109] Rule m: IF X1 is A 1m …and X i is A im THAN T m isβ m m = 1, 2, ..., K

[0110] Among them, X i Let A be the i-th dimension of the input variable. im T is the membership function of the m-th rule with respect to the i-th input variable. m It is the output of the m-th rule, β m Let A be the consequent parameter of the m-th rule, and K be the number of fuzzy rules. The constructed membership function A... im As shown below:

[0111]

[0112] Among them, c im and σ im These represent the center and width of the membership function, respectively.

[0113] The output of the fuzzy rule layer is represented as follows:

[0114]

[0115] The above formula shows that the fuzzy rule layer contains K fuzzy rules.

[0116] Secondly, a cascaded hierarchical incremental learning model with a three-layer cascaded structure is established. In the three-layer cascaded structure, feature data subsets divided based on importance evaluation indicators are fed into each level of the model in batches.

[0117] The learning process of the first-level substructure can be represented as follows:

[0118]

[0119] In the formula, g1 is the Sigmoid activation function, and h1 is the fuzzy rule construction process, containing K1 fuzzy rules. and The input weights are randomly generated. and The input bias is randomly generated, and L1 is the number of neurons used. The learning result of the feature mapping node part is... The result of incremental learning is Together, they constitute the output of the first-layer substructure. y1 is and The result of the second-dimensional concatenation contains a total of 2L1 neuron nodes. Furthermore, and The values ​​are generated from random numbers uniformly distributed on [-λ1, λ1], where λ1 = [0.5, 250]. To ensure the approximation performance of the model, the configuration process of w1 and b1 is constrained by the following hierarchical supervision mechanism:

[0120]

[0121] in, yes The elements in the l1-th column, l1 = 1, 2, ..., L1, as well as Select The corresponding w1 and b1 serve as the input bias and weight of this substructure. When l1 = L1, the configuration process of w1 and b1 is completed.

[0122] The learning process of the second-layer substructure is represented as follows:

[0123]

[0124] In the formula, g2 is the Sigmoid activation function, and h2 is the fuzzy rule construction process, containing K2 fuzzy rules. and The input weights are randomly generated. and The input bias is randomly generated, and L2 is the number of neurons used. The learning result of the feature mapping node part is... and The result of incremental learning is The output of the second layer was determined together. It contains a total of 3L2 neuron nodes. Furthermore, and Generated from random numbers uniformly distributed on [-λ2, λ2], where λ2 = [0.5, 250]. The configuration process of w2 and b2 is constrained based on the following hierarchical supervision mechanism:

[0125]

[0126] in, yes The element in column l2, where l2 = 1, 2, ..., L2, as well as Select The corresponding w2 and b2 groups are the input biases and weights of the cost layer substructure. When l2 = L2, the configuration process of w2 and b2 is completed.

[0127] The learning process of the third-layer substructure can be represented as:

[0128]

[0129] In the formula, g3 is the Sigmoid activation function, and h3 is the fuzzy rule construction process, containing K3 rules. and The input weights are randomly generated. and L3 is the randomly generated input bias, and L3 is the number of neurons used. The learning result of the feature mapping node part is... and The result of incremental learning is The output of the third layer was determined together. It contains a total of 3L3 neuron nodes. Furthermore, and The values ​​are generated from random numbers uniformly distributed on [-λ3, λ3], where λ3 = [0.5, 250]. The configuration process of w3 and b3 is constrained through the following hierarchical supervision strategy to ensure the convergence performance of the model:

[0130]

[0131] in, yes The element in column l3, where l3 = 1, 2, ..., L3, as well as Select The corresponding w3 and b3 form the input bias and weight of the cost layer substructure. When l3 = L3, the configuration process of w3 and b3 is completed.

[0132] This constructs the hidden layer of the incremental hierarchical fuzzy neural network, containing L = 2L1 + 3(L2 + L3) neurons and K = K1 + K2 + k3 fuzzy rules. y3 is used as the output of the hidden layer and re-represented as... Its dimensions are N×L.

[0133] Furthermore, based on the hidden layer output of the incremental hierarchical fuzzy neural network obtained by S3, a transferable incremental hierarchical fuzzy neural network training model is obtained through the consequent learning process of the neural network, as specifically described below:

[0134] Based on the output of the hidden layer obtained in S4 Solve for the output and output weights of the incremental hierarchical fuzzy neural network, and design the following objective loss function:

[0135]

[0136] In the formula, H is the output of the hidden layer of the incremental hierarchical fuzzy neural network, β is the output weight, T is the target object, γ1 is the Lasso regularization tuning parameter, γ2 is the Ridge regularization tuning parameter, ||.||1 represents the Manhattan distance, and ||.||2 represents the Euclidean distance. The Elastic-Net regularization strategy is composed of two different regularization methods.

[0137] The output weight β of the incremental hierarchical fuzzy neural network can be expressed as follows:

[0138] β=(H T H+γ2I) + (H T T+γ1I)

[0139] In the formula, (·) + This represents the pseudo-inverse operation, where I is the identity matrix.

[0140] Based on the above, the output of the incremental hierarchical fuzzy neural network... It can be represented as follows:

[0141]

[0142] In the formula, H is the hidden layer output matrix of the incremental hierarchical fuzzy neural network, and β is the output weight. If metabolic fatty liver is used as the training dataset, then in It includes predictions for metabolic fatty liver.

[0143] The incremental hierarchical fuzzy neural network designed is mainly used for metabolic fatty liver data samples to achieve auxiliary diagnosis of metabolic fatty liver by performing classification tasks. Therefore, RMSE, ACC and AUC are used as performance indicators to evaluate the performance of the model.

[0144] The metabolic fatty liver detection model constructed through the above steps can establish a transferable incremental hierarchical fuzzy neural network learning model based on existing datasets. Further biochemical indicators of metabolic fatty liver disease can then be collected to detect whether the patient has metabolic fatty liver disease. The dataset undergoes processing S1-S4 to obtain the transferable incremental hierarchical fuzzy neural network learning model, which includes the feature ranking results and ranking rules of the metabolic fatty liver dataset, the input weights and input biases of each hierarchical structure, and the output weight β obtained from the training process on the training set. The detection result for metabolic fatty liver disease can be obtained by calculating the output of the incremental hierarchical fuzzy neural network.

[0145] Please refer to the detailed process. Figure 3 This figure is a schematic flowchart of a method for analyzing metabolic fatty liver data and detecting the disease, provided by an embodiment of the present invention.

[0146] This invention provides an algorithmic device for an auxiliary diagnostic method for metabolic fatty liver disease. Please refer to... Figure 4 This figure is a schematic block diagram of an apparatus for analyzing and modeling characteristic data of metabolic fatty liver disease, provided by an embodiment of the present invention. The apparatus includes:

[0147] The acquisition unit 410 is used to acquire data and normalize the data to form a sample of metabolic fatty liver characteristic data.

[0148] The first construction unit 420 is used to perform feature analysis based on the metabolic fatty liver disease data sample combined with a probability model, and to divide the data sample into multiple feature data subsets according to the correlation analysis structure between data features.

[0149] The second construction unit 430 is used to determine the cascaded hierarchical model structure and input samples at each level of the incremental hierarchical fuzzy neural network based on the feature data subset partitioning results.

[0150] The third building unit 440 is used to construct the hierarchical model substructure to form the hidden layer of the incremental hierarchical fuzzy neural network.

[0151] The computing unit 450 performs parameter learning based on the cascaded hierarchical fuzzy neural network model and incremental learning strategy, and finally obtains the input layer weights, output weights and output sequence of the incremental hierarchical fuzzy neural network, and obtains the predicted output sequence, thereby obtaining the auxiliary diagnostic results of metabolic fatty liver disease.

[0152] The various embodiments of the present invention are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other. The focus of each embodiment is to highlight the differences from other embodiments. In particular, the system embodiments are relatively brief because they are basically similar to the method embodiments, and relevant information can be found in the detailed description of the method embodiments.

[0153] It should be clarified that the above description represents only embodiments of the present invention and does not limit the invention. For those skilled in the art, the present invention has many possibilities for modification and variation. Any modifications, equivalent substitutions, improvements, etc., that conform to the spirit and principles of the present invention should be included within the scope of the claims of the present invention.

Claims

1. A hierarchical fuzzy neural network incremental modeling method for classifying metabolic fatty liver data, characterized in that, The method includes the following steps: S1. Collect data and normalize the data to form a data sample set of metabolic fatty liver disease; based on the biochemical index detection data associated with metabolic fatty liver disease and other numerical indicators of patients, select the original indicators as the main characteristics describing metabolic fatty liver disease. S2. Based on the metabolic fatty liver data sample set, feature importance analysis is performed using importance evaluation indicators. Based on the analysis results, the data samples are divided into multiple feature data subsets. S3. Based on the feature data subset partitioning results, adaptively determine the number of cascaded hierarchical model structures of the incremental hierarchical fuzzy neural network, and select feature data subsets input into each level of the model; S4. Based on the fuzzy neural network modeling strategy and incremental learning mechanism, construct the base models at each level in the hierarchical model structure to form a cascaded hierarchical incremental learning model. S4 specifically includes: First, a fuzzy neural network is constructed to establish the substructure of the hierarchical network model. In this fuzzy neural network, the first... The fuzzy rule is represented as follows: Rule : IF is and is THAN is , in, For the first Dimensional input variables, It is the first The rule applies to the first... Membership functions of input variables, It is the first The output of each rule For the first The consequent parameter of a rule, The number of fuzzy rules; the constructed membership function. As shown below: in, and These are the center and width of the membership function, respectively; The output of fuzzy regularization is represented as: The above formula shows that the fuzzy rule layer contains A fuzzy rule; Secondly, a cascaded hierarchical incremental learning model with a three-layer cascaded structure is established; in the three-layer cascaded structure, feature data subsets divided based on importance evaluation indicators are entered into each level of the model in batches; In a three-layer cascaded structure, the learning process of the first-layer substructure can be represented as follows: In the formula, For Sigmod activation function, The process of constructing fuzzy rules includes A fuzzy rule, and The input weights are randomly generated. and It is a randomly generated input bias. This refers to the number of neurons used; the learning result for the feature mapping node part is... The result of incremental learning is Together, they constitute the output of the first-layer substructure. , yes and The result of the second-dimensional stitching includes a total of The number of neurons; in addition... and By conforming Generating uniformly distributed random numbers, where, To ensure the convergence performance of the model, and The configuration process is subject to the following hierarchical supervision mechanisms: in, yes The Middle Column elements, , ,as well as Select corresponding and As the input bias and weights of this substructure, when = hour, and The configuration process is complete; The learning process of the second-layer substructure is represented as follows: In the formula, For Sigmod activation function, The process of constructing fuzzy rules includes A fuzzy rule, , and The input weights are randomly generated. , and It is a randomly generated input bias. This refers to the number of neurons used; the learning result for the feature mapping node part is... and The result of incremental learning is Together, they determined the output of the second layer. , contains The number of neurons; in addition... and By conforming Generating uniformly distributed random numbers, where, Based on the following hierarchical supervision mechanism, constraints are formed. and Configuration process: in, yes The Middle Column elements, , ,as well as Select corresponding and As the input bias and weights of this substructure, when = hour, and The configuration process is complete; The learning process of the third-layer substructure is represented as follows: In the formula, For Sigmod activation function, The process of constructing fuzzy rules includes Rule 1 , and The input weights are randomly generated. , and It is a randomly generated input bias. This refers to the number of neurons used; the learning result of the feature mapping node part is... and The result of incremental learning is Together, they determined the output of the third layer. , contains Number of neurons; As the output of the hidden layer of the incremental hierarchical fuzzy neural network, it is re-represented as Its dimensions are , ;also, and By conforming Generating uniformly distributed random numbers, where, ;Through the following hierarchical supervision strategy, to and The configuration process is constrained to ensure the convergence performance of the model: in, yes The Middle Column elements, , ,as well as Select corresponding and As the input bias and weights of this substructure, when = hour, and The configuration process is complete; S5. Based on the cascaded hierarchical incremental learning model, perform parameter learning to obtain a transferable incremental hierarchical fuzzy neural network model, and then carry out data classification of metabolic fatty liver disease.

2. The method according to claim 1, characterized in that, S1 specifically includes: A data sample set is formed by collecting biochemical indicator test data related to metabolic fatty liver disease. The input sequence of the data sample set is represented as follows: It contains characteristic information describing metabolic fatty liver disease, and the output sequence is represented as follows: It contains labeling information describing the diagnosis of metabolic fatty liver disease; For the sample size, For characteristic number, For the number of tags; Map each data point in the data sample set to... To generate normalized data samples, the normalization formula is as follows: In the formula, It is the first in the data sample set line, number j The original data points of the column, It is the first in the data sample set The minimum value of the feature to which the column belongs. It is the first in the data sample set The maximum value of the feature to which the column belongs. , .

3. The method according to claim 1 or 2, characterized in that, S2 specifically includes: Based on the metabolic fatty liver data sample set described in S1, and combined with the following probability model, an importance evaluation index is established: In the formula, For the data sample Importance evaluation metrics for column feature data For the data sample set, It is the first in the data sample set j Column feature information, For tag information, , Represents Euclidean distance; Based on the aforementioned importance evaluation indicators, the importance of each column of feature data in the data sample is calculated, and the data is sorted in descending order according to the importance evaluation indicators. This completes the data feature importance analysis results, dividing the data sample into... The feature data subset is the feature columns corresponding to the top 30% of the data. The feature columns corresponding to the next 30% are used as a subset of feature data. The feature columns corresponding to the last 40% are used as a subset of feature data. , For the sample size, The total characteristic number, , and This represents the number of features contained in a subset of samples, and Based on the established importance evaluation indicators, the feature data subset , and The importance of the representative features decreases in that order.

4. The method according to claim 1 or 2, characterized in that, S3 specifically includes: Based on the partitioning results of the metabolic fatty liver disease dataset, the original data sample set was divided into three feature data subsets after feature sorting. , and A hierarchical network with a three-layer cascade structure is represented as follows: That is: In the formula, This represents the output of the first layer in a three-layer cascaded structure, whose input is a subset of feature data. , Represents the first substructure; This represents the output of the second layer in a three-layer cascaded structure, whose input is a subset of feature data. Compared with the output of the first substructure, Represents the second substructure; This represents the output of the third layer in a three-layer cascaded structure, which is also the output of the last layer. Its input is a subset of feature data. With the output of the second substructure, This represents the third substructure.

5. The method according to claim 3, characterized in that, S3 specifically includes: Based on the partitioning results of the metabolic fatty liver disease dataset, the original data sample set was divided into three feature data subsets after feature sorting. , and A hierarchical network with a three-layer cascade structure is represented as follows: That is: In the formula, This represents the output of the first layer in a three-layer cascaded structure, whose input is a subset of feature data. , Represents the first substructure; This represents the output of the second layer in a three-layer cascaded structure, whose input is a subset of feature data. Compared with the output of the first substructure, Represents the second substructure; This represents the output of the third layer in a three-layer cascaded structure, which is also the output of the last layer. Its input is a subset of feature data. With the output of the second substructure, This represents the third substructure.

6. The method according to claim 1, 2, or 5, characterized in that, S5 specifically includes: Based on the output of the hidden layer obtained in S4 Solve for the output and output weights of the incremental hierarchical fuzzy neural network, and design the following objective loss function: In the formula, To incrementally optimize the hidden layer output of the hierarchical fuzzy neural network, To output weights, For the target object, These are Lasso regularized tuning parameters. These are the Ridge regularization tuning parameters. Represents Manhattan distance, Representing Euclidean distance, the Elastic-Net regularization strategy is composed of two different regularization methods. Output weights of incremental hierarchical fuzzy neural network The statement is as follows: In particular, In the formula, Represents pseudo-inverse operation. It is the identity matrix; Based on the above, the output of the incremental hierarchical fuzzy neural network... It is expressed as follows: In the formula, To generate the hidden layer output matrix of the incremental hierarchical fuzzy neural network, For output weights.

7. An apparatus for incremental modeling of hierarchical fuzzy neural networks for metabolic fatty liver data classification as described in claim 1, characterized in that, include: The data acquisition unit is used to collect data and normalize the data to form a sample of metabolic fatty liver characteristic data. The first construction unit is used to perform feature analysis based on the metabolic fatty liver disease data samples combined with a probability model, and to divide the data samples according to the correlation analysis structure between data features to form multiple feature data subsets. The second construction unit is used to determine the cascaded hierarchical model structure of the incremental hierarchical fuzzy neural network and the input samples at each level based on the feature data subset partitioning results; The third building unit is used to construct hierarchical model substructures, forming the hidden layers of the incremental hierarchical fuzzy neural network; The computational unit learns parameters based on the cascaded hierarchical fuzzy neural network model and incremental learning strategy, and finally obtains the input layer weights, output weights and output sequence of the incremental hierarchical fuzzy neural network, and uses this to obtain the predicted output sequence, thereby obtaining the prediction result of metabolic fatty liver.

Images