Multi-classification auxiliary prediction method and system for metabolism-related fatty liver diseases, electronic equipment and storage medium

Abstract

The invention discloses a multi-classification auxiliary prediction method and system for metabolism-related fatty liver diseases, electronic equipment and a storage medium, belongs to the field of medical data processing, and is used for solving the problem that a traditional MAFLD binary prediction model cannot fully reflect continuous pedigree of disease severity. The method comprises the following steps: acquiring clinical data related to classification prediction of metabolism-related fatty liver diseases; performing label coding, missing value filling and feature screening on the clinical data; and inputting the screened features into a TabNet classification model optimized by a particle swarm algorithm, and outputting a risk classification result of the metabolism-related fatty liver disease. According to the method, the severity of the fatty liver can be more accurately evaluated, the liver health condition of a patient is comprehensively reflected, and more accurate treatment guidance is provided for clinical diagnosis.

Description

Technical Field

[0001] The present invention relates to medical data processing, and more particularly to a multi-classification auxiliary prediction method and system, electronic equipment, and storage medium for metabolic-related fatty liver disease. Background Art

[0002] Metabolic-associated fatty liver disease (MAFLD), a common chronic liver disease, reflects the pathological process of excessive fat accumulation in the liver. Hepatic steatosis, or fatty liver, is a hallmark of MAFLD. Therefore, early prevention and treatment of MAFLD are crucial. Early detection and management can help patients control or reverse the disease, thereby improving their overall health and well-being, which is crucial to the overall prevention process.

[0003] Currently, the clinical diagnosis of MAFLD relies primarily on imaging studies (such as liver ultrasound and CT) and liver biopsy. However, existing methods have significant limitations. Ultrasound and CT have insufficient sensitivity and limited efficacy in detecting early-stage fatty liver disease, requiring a liver steatosis >30% for identification. Although liver biopsy is the gold standard, its invasive nature can lead to complications and low patient compliance, making it difficult to apply for large-scale early screening. Traditional binary prediction models for MAFLD (e.g., distinguishing between normal and fatty liver states or detecting significant fibrosis / NASH) fail to fully reflect the continuous spectrum of disease severity (e.g., normal, mild, and moderate to severe hepatic steatosis). This oversimplified model limits the ability to refine the risk of disease progression, optimize personalized treatment strategies, and comprehensively evaluate the effectiveness of targeted interventions. In summary, the development of a highly accurate, non-invasive, and population-applicable multi-classification auxiliary diagnostic model for metabolic-related fatty liver disease based on electronic clinical records is crucial for the effective prevention and treatment of metabolic-related fatty liver disease. Summary of the Invention

[0004] An object of the present invention is to solve at least the above problems and to provide at least the advantages which will be described hereinafter.

[0005] Another object of the present invention is to provide a multi-classification auxiliary prediction method and system, electronic device, and storage medium for metabolic-related fatty liver disease. Compared to traditional binary classification models, the multi-classification model in this application can more accurately assess the severity of fatty liver disease, comprehensively reflect the patient's liver health, and provide more accurate treatment guidance for clinical diagnosis. This model facilitates early identification of the disease and timely intervention for at-risk populations, effectively delaying disease progression and enhancing the scientific and practical value of clinical management.

[0006] In order to achieve these purposes and other advantages according to the present invention, a multi-classification auxiliary prediction method for metabolic-related fatty liver disease is provided, comprising the following steps: Obtain clinical data relevant to the classification and prediction of metabolic-related fatty liver disease; performing label encoding, missing value filling, and feature screening on the clinical data, wherein the feature screening utilizes a feature screening method combining an expression tree and a genetic algorithm to process the clinical data, thereby obtaining features with a high-rank correlation with metabolism-related fatty liver disease; The filtered features are input into the TabNet classification model optimized by the particle swarm algorithm to output the risk classification results of metabolic-related fatty liver disease. The classification includes three categories: non-metabolism-related fatty liver disease, mild metabolism-related fatty liver disease, and moderate to severe metabolism-related fatty liver disease. The hyperparameters of the TabNet classification model are determined by particle swarm algorithm optimization, and the optimization goal is to maximize the three-fold cross-validation balanced accuracy; The specific process of building the TabNet classification model optimized by the particle swarm algorithm includes: The stratified sampling method was used to divide the filtered feature data into training set and test set in an 8:2 ratio; Set the search space: the dimension n of the feature transformation in the decision step d ∈[8,16], the feature dimension n of each layer in the attention mechanism a ∈[8,16], number of decision steps n steps ∈[3,5], attention sparse update rate gamma∈[1.0,2.0], sparse regularization coefficient lambda sparse ∈[0.0001, 0.01], learning rate ∈[0.001, 0.1]; A 10-particle swarm was used to iterate 10 times in 6-dimensional space, and the negative value of the three-fold cross-validation equilibrium accuracy was used as the objective function; Output optimal parameter combination: n d =16, n a =16, n steps =3, gamma=1.3, lambda sparse =0.001, learning rate = 0.02; The TabNet classification model built with the optimal parameter combination is trained, with the maximum number of training rounds set to 100, EarlyStopping=20, and sparse attention mechanism.

[0007] Preferably, the process of processing the clinical data using the feature screening method includes: Construct an expression tree, whose leaf nodes are the features in the clinical data and whose internal nodes are the operators in {+, -, *, / , log, sqrt}; The expression tree structure is iteratively optimized through the selection, crossover and mutation operations of the genetic algorithm, and new features are screened and generated using the Spearman rank correlation coefficient as the fitness function.

[0008] Preferably, the process of optimizing the expression tree by the genetic algorithm includes: Limit the maximum depth of the tree when initializing the population; The Spearman rank correlation coefficient between the new features generated by the expression tree and the classification labels of metabolic-related fatty liver disease is used as the individual fitness; The individual with the highest fitness is retained in each generation, and the 11 features with the highest correlation coefficients are selected after 30 generations of iteration.

[0009] Preferably, the label coding refers to label coding of categorical variables in clinical data, including coding of gender variables and coding of outcome variables. The coding rule for gender variables is that female is coded as 0 and male is coded as 1. The coding rule for outcome variables is that non-metabolism-related fatty liver disease is 0, mild metabolism-related fatty liver disease is 1, and moderate to severe metabolism-related fatty liver disease is 2.

[0010] Preferably, the missing value filling adopts KNN algorithm.

[0011] Preferably, before feature screening, non-outcome variable data in the clinical data are subjected to Z-score standardization to eliminate feature dimension differences.

[0012] Preferably, the clinical data include: age, sex, height, weight, waist circumference, systolic blood pressure, diastolic blood pressure, body mass index, neutrophil count, neutrophil ratio, lymphocyte count, lymphocyte percentage, monocyte count, red blood cell count, red blood cell distribution width cv, red blood cell distribution width SD, platelets, total bilirubin, direct bilirubin, indirect bilirubin, glucose, total cholesterol, triglycerides, high-density lipoprotein, low-density lipoprotein, glycated hemoglobin, uric acid, and serum homocysteine.

[0013] The present invention also provides a multi-classification auxiliary prediction system for metabolic-related fatty liver disease, comprising: A clinical data acquisition module, which is used to acquire clinical data related to the classification and prediction of metabolism-related fatty liver disease; A data preprocessing module is used to preprocess the clinical data, including label encoding, missing value filling, and feature screening. The feature screening uses a feature screening method combining expression trees and genetic algorithms to process the clinical data, thereby obtaining features with high-rank correlation with metabolism-related fatty liver disease; An outcome prediction module, which is used to input the filtered features into a TabNet classification model optimized by a particle swarm algorithm, and output a risk classification result for metabolic-related fatty liver disease, which includes three categories: non-metabolism-related fatty liver disease, mild metabolic-related fatty liver disease, and moderate to severe metabolic-related fatty liver disease. The hyperparameters of the TabNet classification model are determined by particle swarm algorithm optimization, with the optimization goal of maximizing the three-fold cross-validation equilibrium accuracy; The specific process of building the TabNet classification model optimized by the particle swarm algorithm includes: The stratified sampling method was used to divide the filtered feature data into training set and test set in an 8:2 ratio; Set the search space: the dimension n of the feature transformation in the decision step d ∈[8,16], the feature dimension n of each layer in the attention mechanism a ∈[8,16], number of decision steps n steps ∈[3,5], attention sparse update rate gamma∈[1.0,2.0], sparse regularization coefficient lambda sparse ∈[0.0001, 0.01], learning rate ∈[0.001, 0.1]; A 10-particle swarm was used to iterate 10 times in 6-dimensional space, and the negative value of the three-fold cross-validation equilibrium accuracy was used as the objective function; Output optimal parameter combination: n d =16, n a =16, n steps =3, gamma=1.3, lambda sparse =0.001, learning rate = 0.02; The TabNet classification model built with the optimal parameter combination is trained, with the maximum number of training rounds set to 100, EarlyStopping=20, and sparse attention mechanism.

[0014] The present invention also provides an electronic device comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the above-mentioned multi-classification auxiliary prediction method for metabolic-related fatty liver disease.

[0015] The present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the multi-classification auxiliary prediction method for metabolism-related fatty liver disease.

[0016] The present invention has at least the following beneficial effects: Traditional binary prediction models for metabolic dysfunction-associated fatty liver disease (MAFLD) (e.g., distinguishing between normal and fatty liver states or detecting significant fibrosis / NASH) fail to fully reflect the continuum of disease severity (e.g., normal, mild, and moderate to severe hepatic steatosis). This oversimplified approach limits the ability to refine the assessment of disease progression risk, optimize personalized treatment strategies, and comprehensively evaluate the effectiveness of targeted interventions.

[0017] This study proposes a non-invasive screening tool for MAFLD. A total of 29,352 subjects were enrolled in the study, and the results were categorized into three categories. Compared to traditional binary classification methods, this tool more comprehensively reflects the continuous gradient of MAFLD severity. Among the subjects, 17,053 were non-MAFLD patients, 7,497 were mild MAFLD patients, and 4,802 were moderate to severe MAFLD patients. Using expression trees and genetic algorithms for feature construction and screening, the method aims to automatically generate new features with high-rank correlations with the target variable from the original dataset. Eleven of the most important risk factors were identified from an initial set of 28 variables. Based on these variables, hyperparameters of the TabNet model were adjusted using PSO. This model enables more accurate assessment of fatty liver severity, comprehensively reflects the patient's liver health, and provides more precise treatment guidance for clinical diagnosis. This model facilitates early identification and intervention of potential risks, effectively delaying disease progression.

[0018] Based on the multi-classification auxiliary prediction method for metabolic-related fatty liver disease proposed in this application, the applicant has constructed a prediction model for the risk of metabolic-related fatty liver disease and developed and launched a clinical auxiliary diagnostic tool. The tool is deployed in the form of a web page, which is easy to operate and responds quickly. It can support clinical medical staff to quickly obtain individualized risk assessment results for patients during actual diagnosis and treatment. This tool can be used as a rapid screening method to assess the risk of MAFLD and provide an important reference for the formulation of personalized treatment plans. It has been piloted in hospitals in Beijing, Xiongan New Area, Shijiazhuang and other places, and its clinical feasibility has been verified. Preliminary patient screening and scenario testing have been carried out to verify the feasibility and practicality of this tool in real clinical scenarios.

[0019] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flow chart of the multi-classification auxiliary prediction method for metabolic-related fatty liver disease according to the present invention; Figure 2The figure is a ROC curve diagram of MAFLD prediction using the TabNet classification model based on PSO optimization described in the present invention; Figure 3 This is a calibration curve diagram for MAFLD prediction using the TabNet classification model based on PSO optimization described in the present invention. DETAILED DESCRIPTION

[0021] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0022] like Figure 1 As shown, the present invention provides a multi-classification auxiliary prediction method for metabolic-related fatty liver disease, comprising the following steps: S1. Obtain clinical data related to the classification and prediction of metabolic-related fatty liver disease; This study employed a retrospective cohort design, using data from 36,563 subjects enrolled at the Health Examination Center of Handan Central Hospital in 2019. Data collection adhered to the principles of the Declaration of Helsinki. All examination information was anonymized to exclude patient privacy identifiers.

[0023] Inclusion criteria: (1) Age 18-100 years old; (2) Complete demographic, anthropometric, blood biochemical index and liver ultrasound records.

[0024] Exclusion criteria: (1) Age <18 or >100 years old; (2) Ultrasound examination results of fatty liver are missing.

[0025] Clinical data related to the classification and prediction of metabolic-related fatty liver disease include age (Age), sex (Sex), height (Height), weight (Weight), waist circumference (WC), systolic blood pressure (SBP), diastolic blood pressure (DBP), body mass index (BMI), neutrophil count (NEUT), neutrophil ratio (NEUT%), lymphocyte count (LY), lymphocyte percentage (LY%), monocyte count (Mono), red blood cell count (RBC), red blood cell distribution width cv (RDW-CV), red blood cell distribution width SD (RDW-SD), platelets (PLT), total bilirubin (TBIL), direct bilirubin (DBIL), indirect bilirubin (IBIL), glucose (GLU), total cholesterol (TCH), triglycerides (TG), high-density lipoprotein (HDL-C), low-density lipoprotein (LDL-C), glycosylated hemoglobin (HbA1c), uric acid (UA), and serum homocysteine ​​(Hcy).

[0026] S2. performing label encoding, missing value filling, and feature screening on the clinical data, wherein the feature screening uses a feature screening method combining an expression tree and a genetic algorithm to process the clinical data, thereby obtaining features with a high-rank correlation with metabolism-related fatty liver disease; Label encoding refers to label encoding of categorical variables in clinical data. For example, encoding of gender variables and outcome variables: gender is encoded as 0 for female and 1 for male; outcome variables are encoded as 0 for non-metabolism-related fatty liver disease, 1 for mild metabolism-related fatty liver disease, and 2 for moderate to severe metabolism-related fatty liver disease.

[0027] Missing value filling refers to the supplementary processing of clinical data that meets the inclusion criteria but is not fully recorded for various reasons, so as to reduce the impact of missing data on research results.

[0028] The missing value filling adopts the KNN algorithm, which specifically includes: 1) Detect missing values ​​in the dataset and mark them as specific identifiers. Since the categorical variables have been labeled, there is no need to convert them into numerical values. 2) Divide the original dataset into a training set and a test set, and randomly generate a certain proportion of missing values ​​(e.g., 10%) in the test set to simulate a real-world scenario; 3) Determine the K value (i.e., the number of nearest neighbors). The default value is usually 5, but it needs to be adjusted based on the characteristics of the dataset. Smaller K values ​​may lead to overfitting, while larger K values ​​may introduce noise. 4) Calculate the distance between samples (usually Euclidean distance) and weight them according to the distance or evenly distribute the weights. Closer neighbors have higher weights and a greater impact on the fill value. 5) For each missing value, select the non-missing values ​​of its K nearest neighbors and calculate the weighted mean or mode as the fill value. The mean is usually used for numerical variables, and the mode is used for categorical variables. 6) Evaluate the effect by comparing the difference between the filled value and the true value (such as MSE, RMSE), and visualize the distribution of missing values ​​and the filling results; 7) Try different K values ​​or distance metrics and choose the configuration that performs best. 8) Save the padded data in a structured format (such as CSV) for subsequent analysis or modeling. Be sure to check that the statistical properties of the padded data are reasonable.

[0029] The feature screening method aims to automatically generate features with high-rank correlations with metabolic-related fatty liver disease from the original dataset. The feature screening method uses a feature screening method that combines expression trees and genetic algorithms to process the clinical data. The method uses the Spearman rank correlation coefficient as the feature evaluation criterion, models the feature construction problem as a combinatorial optimization problem of symbolic expressions, and uses genetic algorithms to simulate the biological evolution process. The constructed expressions are iteratively optimized to automatically discover the feature expressions with the strongest rank correlation with the target variable. This method can effectively address feature engineering problems in high-dimensional, nonlinear, and complex feature spaces.

[0030] The process of processing the clinical data using the feature screening method includes: a) Read the clinical data dataset (e.g., Excel file) and standardize all non-outcome variable features to eliminate interference from feature dimensionality differences. The standardization method uses the Z-score standardization strategy.

[0031] b) Build an expression tree.

[0032] In the algorithm, each "individual" in the population is defined as a binary expression tree. An individual is an expression tree, a single candidate solution in the population, corresponding to a method for generating new features. The population is a collection of individuals (expression trees) and is the set of all candidate feature expressions in the current generation, that is, all possible new feature combinations being considered by the algorithm. The algorithm searches in parallel, and the population acts like a "net" cast in the feature space, simultaneously exploring many possible solutions.

[0033] The leaf node represents a basic feature in the dataset (age (Age), sex (Sex), height (Height), weight (Weight), waist circumference (WC), systolic blood pressure (SBP), diastolic blood pressure (DBP), body mass index (BMI), neutrophil count (NEUT), neutrophil ratio (NEUT%), lymphocyte count (LY), lymphocyte percentage (LY%), monocyte count (Mono), red blood cell count (RBC), red blood cell distribution width cv (RDW-CV), red blood cell distribution width SD (RDW-SD), platelets (PLT), total bilirubin (TBIL), direct bilirubin (DBIL), indirect bilirubin (IBIL), glucose (GLU), total cholesterol (TCH), triglycerides (TG), high-density lipoprotein (HDL-C), low-density lipoprotein (LDL-C), glycosylated hemoglobin (HbA1c), uric acid (UA), serum homocysteine ​​(Hcy)); Internal nodes represent operators, selected from the set { +, -, *, / , log, sqrt}, corresponding to common mathematical operations such as addition (+), subtraction (-), multiplication (×), division ( / ), logarithm (log), square root (sqrt), etc.

[0034] The expression tree is used as the representation structure of feature construction, and each node represents a feature name or operator.

[0035] Each expression tree is equivalent to a mathematical expression, which is used to combine multiple original features into a new derived feature; c) The expression tree structure is iteratively optimized through the selection, crossover and mutation operations of the genetic algorithm, and new features are screened and generated using the Spearman rank correlation coefficient as the fitness function.

[0036] The process of optimizing an expression tree using a genetic algorithm includes: c1) Population initialization Several expression trees are randomly generated as the first generation. The maximum depth of the tree is limited to control the complexity of the expression. Each tree is composed of a combination of original features and operators.

[0037] c2) Fitness function design For each new feature generated by the expression tree, calculate the Spearman rank correlation coefficient between it and the target variable (three-category MAFLD label). This coefficient serves as the individual fitness, measuring the nonlinear monotonicity of its relationship to the outcome variable. A higher fitness indicates a stronger discriminative ability of the feature for disease classification.

[0038] c3) Evolutionary process Initially, randomly generated expression tree individuals form a population, which is optimized through multiple generations of evolution.

[0039] Selection: Eliminate the individuals with the lowest fitness and retain the individuals with high fitness.

[0040] Crossover: Perform a subtree swap operation on two expression trees to generate new expression combinations.

[0041] Mutation: Randomly replace a node (operator or feature) in the expression tree to enhance diversity.

[0042] Elite retention: The individuals with the highest fitness are retained in each generation to ensure that the optimal solution is not lost.

[0043] c4) Iterative Optimization The above evolution process is repeated (30 generations), and the current optimal expression tree and its fitness are recorded in each generation.

[0044] c5) Multi-feature construction Repeat the above evolution process several times to obtain multiple expression trees as candidate features.

[0045] c6) Final feature screening The Spearman rank correlation coefficients between all constructed expressions and the outcome variable were calculated, and the 11 characteristic variables with the highest correlation were selected as the final modeling inputs. These included sex (Sex), age (Age), diastolic blood pressure (DBP), body mass index (BMI), red blood cell count (RBC), red blood cell distribution width cv (RDW-CV), red blood cell distribution width SD (RDW-SD), glucose (GLU), triglycerides (TG), high-density lipoprotein (HDL-C), and low-density lipoprotein (LDL-C).

[0046] S3. Input the filtered features into the TabNet classification model optimized by the particle swarm algorithm to output the risk classification results of metabolic-related fatty liver disease, which include three categories: non-metabolism-related fatty liver disease, mild metabolic-related fatty liver disease, and moderate to severe metabolic-related fatty liver disease. The hyperparameters of the TabNet classification model are determined by particle swarm algorithm optimization, and the optimization goal is to maximize the three-fold cross-validation balanced accuracy. The specific process of building the TabNet classification model optimized by the particle swarm algorithm includes: i) Using the stratified sampling method, the filtered feature data were divided into training set and test set in an 8:2 ratio; ii) Build a Tabnet benchmark model and define the Tabnet hyperparameters to be optimized and their search space; TabNet was selected as the benchmark model for structured data classification tasks, with cross-validation accuracy used as the evaluation metric to guide hyperparameter optimization. TabNet is a deep neural network architecture designed specifically for tabular data. It combines the interpretability of decision tree models with the feature extraction capabilities of deep learning models, offering excellent modeling performance and interpretability. To improve TabNet's performance in structured data classification tasks, a particle swarm optimization (PSO) algorithm was introduced to automate hyperparameter search for the TabNet model.

[0047] The key parameter settings for Tabnet model construction are as follows: the dimension n of feature transformation in the decision step d ∈[8,16], the feature dimension n of each layer in the attention mechanism a ∈[8,16], number of decision steps n steps ∈[3,5], attention sparse update rate gamma∈[1.0,2.0], sparse regularization coefficient lambda sparse∈[0.0001, 0.01], learning rate ∈[0.001, 0.1]; To automatically find the optimal hyperparameter combination, a PSO algorithm is used for global search. This algorithm simulates the foraging behavior of a flock of birds and optimizes parameters through collaborative search. The goal is to maximize the model's three-fold cross-validation Balanced Accuracy. For each candidate hyperparameter combination, the negative of this score is returned as the PSO minimization objective function.

[0048] iii) Initialize and run the PSO optimizer; Use pyswarms.single.GlobalBestPSO to configure the optimizer parameters, number of particles: 10, dimension: 6 (corresponding to 6 hyperparameters), acceleration coefficients (c1, c2) and inertia weight (w) to control the search behavior.

[0049] Set the number of iterations to 10.

[0050] Call the .optimize() method to run the search and record the optimal parameter combination and the corresponding score.

[0051] The optimal parameter combination of the final output is: n d =16, n a =16, n steps =3, gamma=1.3, lambda sparse =0.001, learning rate=0.02; iv) Train the final TabNet model based on the optimal parameters; The model is constructed and trained using the optimal hyperparameter combination obtained by particle swarm search, with a maximum number of training rounds set to 100, EarlyStopping (with a patience value of 20), and a sparse attention mechanism (sparsemax).

[0052] The accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), F1 score, mean bias error (MBE) (see Table 1) and area under the receiver operating characteristic curve (AUC) parameters (see Figure 2 ) to evaluate the model performance.

[0053] Table 1 Tabnet model performance Set AUC MBE Accuracy Kappa F1 score Sensitivity Specificity PPV NPV training set 0.837 0.041 0.679 0.450 0.686 0.614 0.832 0.611 0.823 Test set 0.843 0.030 0.691 0.469 0.698 0.624 0.838 0.624 0.830 The TabNet classification model based on PSO optimization in this application performed stably in the three-class fatty liver prediction task, with a test set AUC of 0.843 ( Figure 2), indicating that the model is stable in distinguishing the three types of fatty liver risk and has good discrimination. The mean bias error (MBE) of the training set is 0.041, and the test set is 0.030, which is close to 0. The prediction deviation is small, indicating that the model is stable. Calibration curve ( Figure 3 ) is close to the diagonal, the model's predictions are highly consistent with the actual situation and have good calibration capabilities.

[0054] Based on the same inventive concept, the present invention also provides a multi-classification auxiliary prediction system for metabolic-related fatty liver disease. The multi-classification auxiliary prediction system can be a personal computer, a server, or other systems that implement the aforementioned multi-classification auxiliary prediction method for metabolic-related fatty liver disease.

[0055] The multi-classification auxiliary prediction system for metabolic-related fatty liver disease provided in this embodiment includes: A clinical data acquisition module, which is used to acquire clinical data related to the classification and prediction of metabolism-related fatty liver disease; A data preprocessing module, which is used to perform label encoding, missing value filling, and feature screening on the clinical data, wherein the feature screening uses a feature screening method combining an expression tree and a genetic algorithm to process the clinical data, thereby obtaining features with high-rank correlation with metabolism-related fatty liver disease; An outcome prediction module, which is used to input the filtered features into a TabNet classification model optimized by a particle swarm algorithm, and output a risk classification result for metabolic-related fatty liver disease, which includes three categories: non-metabolism-related fatty liver disease, mild metabolic-related fatty liver disease, and moderate to severe metabolic-related fatty liver disease. The hyperparameters of the TabNet classification model are determined by particle swarm algorithm optimization, with the optimization goal of maximizing the three-fold cross-validation equilibrium accuracy; The specific process of optimizing the hyperparameters of the TabNet classification model using the particle swarm algorithm includes: The stratified sampling method was used to divide the filtered feature data into training set and test set in an 8:2 ratio; Set the search space: the dimension n of the feature transformation in the decision step d ∈[8,16], the feature dimension n of each layer in the attention mechanism a ∈[8,16], number of decision steps n steps ∈[3,5], attention sparse update rate gamma∈[1.0,2.0], sparse regularization coefficient lambda sparse ∈[0.0001, 0.01], learning rate ∈[0.001, 0.1]; A 10-particle swarm was used to iterate 10 times in 6-dimensional space, and the negative value of the three-fold cross-validation equilibrium accuracy was used as the objective function; Output optimal parameter combination: n d =16, n a =16, n steps =3, gamma=1.3, lambda sparse =0.001, learning rate = 0.02; The TabNet classification model built with the optimal parameter combination is trained, with the maximum number of training rounds set to 100, EarlyStopping=20, and sparse attention mechanism.

[0056] All relevant contents of each step involved in the embodiment of the multi-classification auxiliary prediction method for metabolic-related fatty liver disease can be referred to the functional description of the functional modules corresponding to the multi-classification auxiliary prediction system in the embodiment of the present application, and will not be repeated here.

[0057] The division of modules in the embodiments of the present application is illustrative and is merely a logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional modules in the various embodiments of the present invention may be integrated into a single processor, or may exist physically separately, or two or more modules may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or software functional modules.

[0058] The present invention also provides an electronic device comprising: at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to cause the at least one processor to perform the multi-classification auxiliary prediction method for metabolic-related fatty liver disease. The electronic device can be any terminal device including a mobile phone, a laptop computer, a desktop computer, a tablet computer, a PDA (Personal Digital Assistant), a POS (Point of Sales), an in-vehicle computer, or the like.

[0059] The present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the multi-classification auxiliary prediction method for metabolism-related fatty liver disease.

[0060] Through the description of the above embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware, and of course can also be implemented by dedicated hardware including application-specific integrated circuits, dedicated CPUs, dedicated memories, dedicated components, etc. In general, all functions performed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structures used to implement the same function can also be diverse, such as analog circuits, digital circuits, or dedicated circuits. However, for the present invention, software program implementation is a better implementation method in most cases. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc., and includes a number of instructions for causing a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in each embodiment of the present invention.

[0061] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the description and implementation methods. They can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A multi-classification auxiliary prediction method for metabolic-related fatty liver disease, characterized in that: The following steps are involved: Obtain clinical data relevant to the classification and prediction of metabolic-related fatty liver disease; performing label encoding, missing value filling, and feature screening on the clinical data, wherein the feature screening utilizes a feature screening method combining an expression tree and a genetic algorithm to process the clinical data, thereby obtaining features with a high-rank correlation with metabolism-related fatty liver disease; The filtered features are input into the TabNet classification model optimized by the particle swarm algorithm to output the risk classification results of metabolic-related fatty liver disease. The classification includes three categories: non-metabolism-related fatty liver disease, mild metabolism-related fatty liver disease, and moderate to severe metabolism-related fatty liver disease. The hyperparameters of the TabNet classification model are determined by particle swarm algorithm optimization, and the optimization goal is to maximize the three-fold cross-validation balanced accuracy; The specific process of constructing the TabNet classification model optimized by the particle swarm algorithm includes: The stratified sampling method was used to divide the filtered feature data into training set and test set in an 8:2 ratio; Set the search space: the dimension n of the feature transformation in the decision step d ∈[8,16], the feature dimension n of each layer in the attention mechanism a ∈[8,16], number of decision steps n steps ∈[3,5], attention sparse update rate gamma∈[1.0,2.0], sparse regularization coefficient lambda sparse ∈[0.0001, 0.01], learning rate ∈[0.001, 0.1]; A 10-particle swarm was used to iterate 10 times in 6-dimensional space, and the negative value of the three-fold cross-validation equilibrium accuracy was used as the objective function; Output optimal parameter combination: n d =16, n a =16, n steps =3, gamma=1.3, lambda sparse =0.001, learning rate=0.02; The TabNet classification model built with the optimal parameter combination is trained, with the maximum number of training rounds set to 100, EarlyStopping=20, and sparse attention mechanism.

2. The multi-classification auxiliary prediction method for metabolic-related fatty liver disease according to claim 1, characterized in that: The process of processing the clinical data using the feature screening method includes: Construct an expression tree, whose leaf nodes are the features in the clinical data and whose internal nodes are the operators in {+, -, *, / , log, sqrt}; The expression tree structure is iteratively optimized through the selection, crossover and mutation operations of the genetic algorithm, and new features are screened and generated using the Spearman rank correlation coefficient as the fitness function.

3. The multi-classification auxiliary prediction method for metabolic-related fatty liver disease according to claim 2, characterized in that: The process of optimizing an expression tree using a genetic algorithm includes: Limit the maximum depth of the tree when initializing the population; The Spearman rank correlation coefficient between the new features generated by the expression tree and the classification labels of metabolic-related fatty liver disease is used as the individual fitness; The individual with the highest fitness is retained in each generation, and the 11 features with the highest correlation coefficients are selected after 30 generations of iteration.

4. The multi-classification auxiliary prediction method for metabolic-related fatty liver disease according to claim 1, characterized in that: The label coding refers to label coding of categorical variables in clinical data, including coding of gender variables and coding of outcome variables. The coding rule for gender variables is that female is coded as 0 and male is coded as 1. The coding rule for outcome variables is that non-metabolism-related fatty liver disease is 0, mild metabolism-related fatty liver disease is 1, and moderate to severe metabolism-related fatty liver disease is 2.

5. The multi-classification auxiliary prediction method for metabolic-related fatty liver disease according to claim 1, characterized in that: The missing value filling adopts KNN algorithm.

6. The multi-classification auxiliary prediction method for metabolic-related fatty liver disease according to claim 1, characterized in that: Before feature screening, the non-outcome variable data in the clinical data were normalized using Z-score to eliminate feature dimension differences.

7. The multi-classification auxiliary prediction method for metabolic-related fatty liver disease according to claim 1, characterized in that: The clinical data include: age, sex, height, weight, waist circumference, systolic blood pressure, diastolic blood pressure, body mass index, neutrophil count, neutrophil ratio, lymphocyte count, lymphocyte percentage, monocyte count, red blood cell count, red blood cell distribution width cv, red blood cell distribution width SD, platelets, total bilirubin, direct bilirubin, indirect bilirubin, glucose, total cholesterol, triglycerides, high-density lipoprotein, low-density lipoprotein, glycated hemoglobin, uric acid, and serum homocysteine.

8. A multi-classification auxiliary prediction system for metabolic-related fatty liver disease, characterized by: include: A clinical data acquisition module, which is used to acquire clinical data related to the classification and prediction of metabolism-related fatty liver disease; A data preprocessing module is used to preprocess the clinical data, including label encoding, missing value filling, and feature screening. The feature screening uses a feature screening method combining expression trees and genetic algorithms to process the clinical data, thereby obtaining features with high-rank correlation with metabolism-related fatty liver disease; An outcome prediction module, which is used to input the filtered features into a TabNet classification model optimized by a particle swarm algorithm, and output a risk classification result for metabolic-related fatty liver disease, which includes three categories: non-metabolism-related fatty liver disease, mild metabolic-related fatty liver disease, and moderate to severe metabolic-related fatty liver disease. The hyperparameters of the TabNet classification model are determined by particle swarm algorithm optimization, with the optimization goal of maximizing the three-fold cross-validation equilibrium accuracy; The specific process of building the TabNet classification model optimized by the particle swarm algorithm includes: The stratified sampling method was used to divide the filtered feature data into training set and test set in an 8:2 ratio; Set the search space: the dimension n of the feature transformation in the decision step d ∈[8,16], the feature dimension n of each layer in the attention mechanism a ∈[8,16], number of decision steps n steps ∈[3,5], attention sparse update rate gamma∈[1.0,2.0], sparse regularization coefficient lambda sparse ∈[0.0001, 0.01], learning rate ∈[0.001, 0.1]; A 10-particle swarm was used to iterate 10 times in 6-dimensional space, and the negative value of the three-fold cross-validation equilibrium accuracy was used as the objective function; Output optimal parameter combination: n d =16, n a =16, n steps =3, gamma=1.3, lambda sparse =0.001, learning rate=0.02; The TabNet classification model built with the optimal parameter combination is trained, with the maximum number of training rounds set to 100, EarlyStopping=20, and sparse attention mechanism.

9. An electronic device, characterized in that: include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to cause the at least one processor to perform the method according to any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images