Rare disease auxiliary diagnosis method based on el-topsis weight decision

Abstract

The application discloses a rare disease auxiliary diagnosis method based on EL-TOPSIS weight decision, which comprises the following steps: acquiring clinical data of rare disease patients, performing data cleaning, missing value filling and standardization pretreatment; adopting mutual information method to perform feature selection and screen high information amount features related to disease subtypes; using a synthetic minority over-sampling technique to perform class balancing processing on the data; constructing a decision matrix based on performance indexes of each classifier on a verification set; adopting an ideal solution similarity degree sorting preference method to calculate relative closeness degrees of each classifier and ideal solutions, which are used as integrated voting weights; constructing a weighted voting integrated model, and outputting a final disease subtype classification result; and evaluating and comparing the model. The application combines multi-classifier integration and multi-criteria decision, effectively improves accuracy, stability and generalization ability of rare disease subtype classification, is suitable for small sample, high dimension and unbalanced clinical data scenes, and has important clinical auxiliary diagnosis value.

Description

Technical Field

[0001] This invention relates to the fields of medical informatics, machine learning and clinical decision support technology, and in particular to a rare disease auxiliary diagnosis method based on weight decision of multi-model classifier ensemble (EL) and ideal solution similarity ranking preference (TOPSIS). Background Technology

[0002] Rare diseases, due to their low incidence, complex clinical manifestations, and high heterogeneity, often suffer from misdiagnosis, missed diagnosis, and delayed confirmation. In multi-subtype or multi-class scenarios, the sample size of different subtypes varies significantly, leading to class imbalance in the training data. Simultaneously, strong correlations and redundancy exist among clinical indicators, easily introducing noise and reducing the model's generalization ability. Existing single machine learning classification models are often sensitive to specific data distributions and lack stability under small sample and imbalanced data conditions, making it difficult to simultaneously achieve accuracy and robustness. Therefore, there is an urgent need for an auxiliary diagnostic method that can handle multidimensional clinical data, possess feature selection and class balancing capabilities, and effectively weight and fuse multiple heterogeneous learners. Summary of the Invention

[0003] This invention aims to address the problems of data scarcity, class imbalance, feature redundancy, and insufficient stability of single models in the existing diagnosis of multiple subtypes of rare diseases, and provides a rare disease auxiliary diagnosis method based on EL-TOPSIS ensemble learning to solve the above problems.

[0004] To achieve the above objectives, this invention proposes a rare disease auxiliary diagnosis method based on EL-TOPSIS weighted decision-making.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A rare disease auxiliary diagnostic method based on EL-TOPSIS weighted decision-making includes the following steps:

[0007] Step 1: Obtain clinical multidimensional indicator data of patients with the target rare disease and construct a training dataset containing sample data of multiple patients with the target rare disease;

[0008] Clinical multidimensional indicator data includes multidimensional clinical data of patients obtained from electronic medical record systems (EMR) or clinical databases, specifically including but not limited to the following types of data:

[0009] General information: including the patient's gender, age, medical history, family history, and past medical records;

[0010] Laboratory testing indicators: liver function tests, urinary copper content, serum ceruloplasmin levels, and other medical chemical indicators;

[0011] Physical signs scoring indicators include the patient's clinical signs (such as jaundice score, neurological symptom score, etc.) and the doctor's clinical assessment;

[0012] Structured clinical examination results: Imaging examination results, gene testing results, MRI, CT and other imaging data are converted into numerical or categorical forms. These data will be input in the form of structured data tables and will be appropriately processed and transformed according to the type of data. The diagnostic label is the disease category or subtype of the target rare disease.

[0013] Step two involves preprocessing the training dataset described in step one. This preprocessing includes at least missing value imputation, outlier / invalid value handling, and numerical feature normalization. The specific preprocessing methods are as follows:

[0014] Missing value imputation: For missing data, imputation is performed using valid data from neighboring samples; specifically, by calculating the similarity between each sample and its neighboring samples, missing values ​​are filled using a weighted average method or other imputation methods.

[0015] Outlier and invalid value removal: By setting thresholds (such as the 3sigma principle), obvious outliers are removed. Invalid values ​​(e.g., values ​​indicating incorrect input) are corrected or deleted.

[0016] Min-Max Normalization: Standardize all numerical features by using a mapping method between minimum and maximum values ​​to normalize all feature values ​​to the [0,1] interval, thereby eliminating the dimensional differences between different features and ensuring that each feature has the same weight during training.

[0017] Step 3: For the training dataset after data preprocessing in Step 2, the mutual information feature selection method is used to calculate the mutual information score between each conditional feature and the diagnostic label, and a feature subset is obtained by filtering according to a threshold. This specifically includes the following steps:

[0018] (31) Calculate the mutual information score between each feature and the target variable to reflect the correlation between the feature and the diagnostic label;

[0019] (32) Based on the scores, a threshold is determined using the quartile method to remove features with low contribution. The remaining features will be retained as the feature set used when training the model.

[0020] This method can effectively reduce feature redundancy, improve the computational efficiency of the model, and enhance the interpretability of the model.

[0021] Step four: Perform class balancing on the training samples corresponding to the feature subset described in step three, and use the SMOTE algorithm to perform synthetic oversampling on the minority classes. The specific steps are as follows:

[0022] (41) Category statistics and determination of target sample number: The number of samples of each rare disease subtype in the training dataset after mutual information screening and normalization is counted, and the six subtypes are denoted as:

[0023] ;

[0024] The corresponding sample size is ,make As the target sample size; for each minority class The number of samples to be synthesized is ;

[0025] (42) Feature space and distance metric: Let each preprocessed patient sample be represented as a d-dimensional feature vector: Each component represents a clinical indicator that has undergone missing value imputation, outlier handling, and min-max normalization. The distance between samples is calculated using the Euclidean distance formula:

[0026] ;

[0027] (43) For each minority class The original sample In the same category Within the sample set, calculate its Euclidean distance according to the above-mentioned Euclidean distance metric. The nearest neighbor sample set:

[0028] ,

[0029] in For example, the preset number of nearest neighbors. ;

[0030] (44) For each minority class sample When a new synthetic sample needs to be generated, it is derived from its nearest neighbor set. Randomly select a neighboring sample A random interpolation coefficient λ, taking values ​​in the interval [0,1], is generated, and a new synthetic sample is generated according to the following linear interpolation formula. :

[0031] ,

[0032] That is, for each feature m, we have:

[0033] ,

[0034] The above generation process is carried out in the normalized feature space to ensure that each dimension of the new sample still falls within the preset value range;

[0035] (45) For each minority class c, repeat steps 43 and 44 to generate synthetic samples for all its original samples in a loop until the cumulative number of synthetic samples for that class reaches the required number, so that the updated number of class samples satisfies: This allows the number of samples from the six rare disease subtypes to achieve an approximately balanced distribution in the training set.

[0036] (46) Merge the synthetic samples of all categories with the original samples to obtain a balanced training sample set:

[0037] ,

[0038] in Indicate category The corresponding set of all synthetic samples, after being balanced, is used for training multiple heterogeneous base classifiers, thereby improving the recall rate of a few subtypes and the overall classification performance.

[0039] Step 5: Train multiple heterogeneous base classifiers based on the training samples balanced in Step 4 to obtain each base classification model;

[0040] The process of training multiple heterogeneous base classifiers based on the balanced training samples to obtain base classification models including at least LightGBM, GBDT, Random Forest, SVM, and XGBoost, specifically includes the following steps:

[0041] 1) The training sample set after data preprocessing, mutual information feature filtering, and SMOTE class balancing is denoted as: ,

[0042] in: Indicates the first Feature vectors of a patient sample This indicates the number of feature dimensions retained after filtering. This indicates the rare disease subtype label corresponding to the patient; the subtype labels include six categories: hepatic, neurological, asymptomatic, arthritic, cutaneous, and renal.

[0043] During the training phase, the balanced sample set is divided into a training set and a validation / test set for model fitting and performance evaluation.

[0044] 2) To improve the model's generalization ability and stability in multi-class assisted diagnosis scenarios, multiple classification algorithms with different structures and learning mechanisms were selected as heterogeneous base classifiers to form a base classifier set:

[0045] ,

[0046] These correspond to: ƒ1-LightGBM classifier, ƒ2-GBDT classifier, ƒ3-Random Forest classifier, ƒ4-Support Vector Machine (SVM) classifier, and ƒ5-XGBoost classifier, respectively.

[0047] Heterogeneous models are used because different classifiers have different advantages in nonlinear relationships, feature interactions, multi-class boundaries, minority class recognition, and robustness. They can be integrated after weights are calculated using TOPSIS.

[0048] LightGBM is an ensemble classifier based on gradient boosting decision trees. It employs a histogram-based splitting strategy, discretizing continuous features into multiple buckets to reduce the number of candidate split points, thereby lowering training complexity and improving training efficiency. Its specific training process is as follows:

[0049] 1-1. Initialize model output.

[0050] 1-2. Calculate the gradient of the loss function of the current model on the training samples.

[0051] 1-3. Using the negative gradient direction as the fitting target, train a new decision tree.

[0052] 1-4. Add the decision tree to the existing tree set.

[0053] 1-5. Update the overall model output.

[0054] 1-6. Repeat the above process until the preset number of trees or the convergence condition is reached. If the wheel model is denoted as , then its recursive form is:

[0055] ,

[0056] in: For the first A new tree was added. The learning rate;

[0057] GBDT is a gradient boosting decision tree model, and its specific training steps include:

[0058] 2-1. Initialize the classification model with a constant function.

[0059] 2-2. For each iteration, calculate the residual or negative gradient of the current model on each training sample.

[0060] 2-3. Train a new regression tree / classification tree based on the residuals.

[0061] 2-4. Multiply the new tree by the learning rate and then add it to the current model.

[0062] 2-5. After reaching the preset number of iterations, output the final classifier, whose recursive expression is:

[0063] ,

[0064] in: For the first Wheel integration model, To create a new tree for fitting the residuals, The learning rate;

[0065] Random forests belong to the Bagging ensemble model category. They construct multiple independent decision trees in parallel and merge the results of each tree through voting to reduce the variance problem caused by overfitting of a single tree. The specific training process includes:

[0066] 3-1. Generate multiple Bootstrap sample subsets from the training sample set using sampling with replacement.

[0067] 3-2. Train a decision tree for each Bootstrap sample subset.

[0068] 3-3. When splitting at each node, the optimal splitting variable is selected only from a randomly selected subset of features.

[0069] 3-4. Repeat the above process to generate a predetermined number of decision trees.

[0070] 3-5. Perform majority voting on the output of all trees to obtain the final predicted category. If the first tree... The output of the trees is Then the overall output of the random forest is:

[0071] ,

[0072] in: The number of decision trees; Indicates the majority vote;

[0073] The specific training steps for a Support Vector Machine (SVM) classifier are as follows:

[0074] 4-1. Map the input feature vector to the kernel space.

[0075] 4-2. Construct the maximum margin optimization problem.

[0076] 4-3. Solve for the support vectors and their Lagrange multipliers.

[0077] 4-4. Obtain the optimal classification hyperplane.

[0078] 4-5. For multi-class classification tasks, use a one-to-many or one-to-one strategy to output the classification results for each class.

[0079] 4-6. After enabling probability output, the prediction probability for each category can be further obtained. The SVM optimization objective is:

[0080] ,

[0081] Satisfy constraints:

[0082] ,

[0083] in: For kernel mapping; For regularization parameters; These are slack variables;

[0084] XGBoost is an improved gradient boosting tree model that introduces a regularization term into the traditional boosting tree model and supports parallel computation, tree pruning, and efficient feature splitting search. Its training process includes:

[0085] 5-1. Initialize model predictions.

[0086] 5-2. Calculate the first and second gradients based on the current prediction results.

[0087] 5-3. Select the optimal split point based on the objective function gain.

[0088] 5-4. Generate new trees and prune them.

[0089] 5-5. Add the new tree to the existing model.

[0090] 5-6. Repeat the iteration until the maximum number of trees is reached or the convergence condition is met. The overall objective function is:

[0091] ,

[0092] in: The loss function; For the first Tree; This is a regularization term used to control model complexity;

[0093] 3) After the above five basic classifiers are trained, corresponding basic classification models are formed:

[0094] ,

[0095] For any input patient sample X, each classification model outputs a predicted class, or a probability distribution for the six rare disease subtypes:

[0096] ,

[0097] Where: k represents the k-th base classifier; L represents the total number of classes; and represents the probability that the classifier predicts a sample belongs to the j-th class. Each base classifier selects the class with the highest probability as its single-model output.

[0098] .

[0099] This output will be used in subsequent steps for performance evaluation, decision matrix construction, TOPSIS weight calculation, and weighted voting fusion.

[0100] Step six involves calculating multiple classification metrics for each basic classification model on the validation or test set, and constructing a decision matrix with the classification metrics as attributes. This includes the following steps:

[0101] 1) The multiple heterogeneous basic classification models trained in step six are denoted as follows: Where K represents the number of heterogeneous basic classification models.

[0102] The sample set to be evaluated is denoted as: ,

[0103] Where: is the first Feature vectors of each test sample; This corresponds to the actual category label; the category label is one of the six subtypes of Wilson's disease.

[0104] For any basic classification model Test samples After inputting the data into the model, we obtain its prediction results:

[0105] ,

[0106] Or obtain the corresponding category probability output: ,in: The total number of categories, Indicates the first The classification model predicts the first... The sample belongs to the first The probability of a class;

[0107] After each classification model completes its predictions for all test samples, it generates a corresponding set of prediction results for subsequent performance evaluation.

[0108] 2) For each basic classification model, construct the confusion matrix corresponding to that model based on the true label and the predicted label:

[0109] ,

[0110] in: Indicates the first In the nth model, the true category is the nth Class and predicted as the first The number of samples in each class; For the dimensions of the multi-class confusion matrix;

[0111] 3) After obtaining the confusion matrix of each model, calculate six performance evaluation metrics as attribute columns of the subsequent decision matrix; the six metrics used are: Accuracy, Precision, Recall, F1Score, Specificity, and AUROC (area under the ROC curve).

[0112] The calculation formulas are as follows:

[0113] Accuracy represents the proportion of correctly classified samples out of the total number of test samples, expressed as:

[0114] ,

[0115] in: Indicates the first The number of correctly classified samples of a class; Indicates the total number of test samples;

[0116] Precision, for the first Class, precision is defined as:

[0117] ,

[0118] in: This indicates that it was incorrectly predicted as the first. The number of samples in each class; for multi-class classification problems, a weighted average method is used to obtain the overall precision;

[0119] Recall, for the first Class, recall rate is defined as:

[0120] ,

[0121] in: Indicates that it truly belongs to the first The total recall of a multi-class system can be calculated using a weighted average method, which measures the number of samples that are classified as one class but are predicted as other classes.

[0122] The F1 score is used to comprehensively reflect the balance between precision and recall. kind: The overall F1 score can be obtained by weighted averaging.

[0123] Specificity measures a model's ability to identify negative class samples. The class is defined as:

[0124] ,

[0125] in: Indicates that it does not belong to the first The number of samples that are correctly identified as not belonging to a class; in multi-class scenarios, the overall specificity can be the average of the specificities of each class.

[0126] AUROC (Area Under the ROC Curve) is used to measure the ability of a model to distinguish between different classes. In multi-class classification problems, a one-to-many strategy is used to calculate the area under the ROC curve for each class separately, and then the average value is taken as the overall AUROC.

[0127] 4) For each basic classification model The above six performance metrics on the test set are combined into a performance evaluation vector: ,in: Indicates accuracy; Indicates accuracy; Indicates recall rate; Indicates the F1 score; Indicates specificity; This represents AUROC; that is, each base classifier is ultimately mapped to a six-dimensional performance description vector.

[0128] 5) All The performance evaluation vectors of the base classifiers are arranged in rows to construct the original decision matrix:

[0129] .

[0130] In this implementation: number of rows There are five basic classifiers; the number of columns is 6, corresponding to six performance metrics.

[0131] 6) To reduce the impact of accidental classifications or fluctuations in a single experiment, each basic classification model was repeatedly trained and tested. Six evaluation indicators were recorded in each experiment, and the mean and standard deviation of each indicator were calculated to form more stable evaluation results. The model of the first The indicator in the first The values ​​taken in this experiment are denoted as follows: Then its mean and standard deviation are respectively:

[0132]

[0133] ,

[0134] in: To repeat the experiment a certain number of times, further intervals were used. This indicates the range of fluctuation of the model on this evaluation index, thus reflecting the stability of the model.

[0135] Step 7: Standardize and weight the decision matrix described in Step 6 using the TOPSIS method, determine the positive and negative ideal solutions, and calculate the relative proximity coefficient. This relative proximity coefficient is then used as the voting weight for the corresponding base classification model. After constructing the base classifier performance decision matrix, the TOPSIS method is used to calculate the voting weight for each base classification model. Specifically, this also includes the following steps:

[0136] 1) Assume there is a total There are 1 basic classification model, and the number of evaluation indicators is 1. The original decision matrix is ​​constructed as follows: ,in , indicating the first A basic classification model; , , indicating the first One evaluation indicator;

[0137] 2) Since different evaluation indicators have different dimensions, the decision matrix needs to be dimensionless, using vector normalization (a standard practice in TOPSIS):

[0138] ,

[0139] Obtain the standardized matrix: ;

[0140] 3) Setting indicator weights: The importance of each evaluation indicator to model performance can be set to equal weight, or a weight vector can be set: ,satisfy The weighted normalized matrix is ​​obtained by weighting the normalized matrix: , denoted as: ;

[0141] 4) Determine the positive and negative ideal solutions. Since all evaluation indicators are benefit-type indicators (the higher the value, the better), as follows:

[0142] Positive ideal solution: ,in: This represents the combined vector when each indicator achieves its optimal performance.

[0143] Negative ideal solution: ,in: This represents the combined vector when each indicator reaches its worst performance.

[0144] 5) For each basic classification model Calculate the Euclidean distance between it and the positive and negative ideal solutions.

[0145] Distance from the ideal solution: ,

[0146] Distance to negative ideal solution: ;

[0147] 6) Calculate the relative proximity coefficient. According to the TOPSIS principle, define the first... The relative proximity of the base classifiers is:

[0148] ,

[0149] in: , The larger the value, the closer the classifier is to the ideal solution;

[0150] 7) Compile the relative proximity coefficients of all basic classification models into a weight vector:

[0151] ,

[0152] And it can be further normalized: This yields the classifier weights used for the final weighted voting.

[0153] Step 8: Feature construction of the samples to be diagnosed. The samples to be diagnosed are processed in the same way as the data processing in the training phase. The prediction results are then output by each basic classification model. Weighted voting is performed based on the voting weights to output the diagnosis category or subtype of the rare disease.

[0154] After obtaining the voting weights of each basic classification model, a weighted voting fusion based on weights is performed on the patient samples to obtain the final rare disease diagnosis category or subtype result. The specific steps include the following:

[0155] 1) Construction of features for patients awaiting diagnosis: Obtain multidimensional clinical indicator data of patients awaiting diagnosis, and process them according to the same data processing procedures as in the training phase, including: imputation of missing values, handling of outliers or invalid values, min-max normalization, and mutual information filtering to obtain feature subsets. Let the feature vector of the patients awaiting diagnosis be represented as: ,in: These are the filtered feature dimensions;

[0156] 2) Basic classification model prediction output. Input the samples to be diagnosed into the pre-trained... Basic classification models: The prediction results of each model are obtained:

[0157] ,

[0158] in: In this embodiment, the category or subtype of a rare disease is indicated as follows: These correspond to the liver type, neurological type, asymptomatic type, joint type, skin type, and kidney type, respectively.

[0159] Each basic classification model can also output the probability distribution of belonging to each category:

[0160] ,

[0161] in: ,

[0162] The predicted category of the model can be determined by the maximum probability: ;

[0163] 3) In step eight, the relative proximity coefficients of each basic classification model are calculated using the TOPSIS method: And form the model weight vector: ,in: Or further normalize to: ;

[0164] 4) Construct a category-weighted scoring function for each possible diagnostic category. Calculate its weighted voting score: ,

[0165] in: For indicator functions; when the first The model predicts the following categories: hour, ,otherwise: ,therefore: This represents all basic classification models for categories. The weighted degree of support;

[0166] 5) The final diagnostic category is obtained by selecting the category with the highest weighted score: ,in: The final output is the diagnostic category or subtype;

[0167] 6) Output diagnostic results. The final system output results include: rare disease diagnosis category or subtype, voting results of each model, model weights, and category weighted scores.

[0168] Beneficial effects

[0169] Compared with the prior art, the present invention has at least the following beneficial effects:

[0170] (1) Reduce the redundancy of clinical indicators by mutual information feature selection, and improve the efficiency and interpretability of model training;

[0171] (2) Improve the identification ability of minority subtypes and reduce model bias by using SMOTE class balancing;

[0172] (3) By integrating multiple evaluation indicators into weights through TOPSIS, the complementary advantages of basic learners can be fully utilized;

[0173] (4) A stable integrated output is obtained through weighted voting, which improves diagnostic accuracy and consistency in small sample and multi-subtype scenarios. Attached Figure Description

[0174] Figure 1 This is a schematic diagram of the overall structure of the present invention;

[0175] Figure 2 This is one of the classification performance characteristics of the classifiers in this invention;

[0176] Figure 3 This is one of the classification performance characteristics of the classifiers in this invention;

[0177] Figure 4 This is one of the classification performance characteristics of the classifiers in this invention;

[0178] Figure 5 This is a comparison of the kappa coefficients of various classifiers in this invention. Detailed Implementation

[0179] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0180] In summary, the physical fitness assessment and pulse diagnosis data acquisition and analysis system of this invention mainly consists of two parts: a data acquisition terminal and a data processing and analysis system. The specific implementation method is as follows... Figure 1 As shown, this invention provides an auxiliary diagnostic method for rare disease subtypes based on EL-TOPSIS, comprising the following steps:

[0181] The first step is to obtain and preprocess the multidimensional clinical indicators of patients awaiting diagnosis.

[0182] (1) This invention uses a multidimensional biomedical dataset of clinical cases of Wilson's disease, which collects clinical characteristics of Wilson's disease patients from 30 different provinces and regions across 34 provincial-level administrative units in China. The dataset contains 50 important attributes and records information on 1623 patients. Among them, 73.99% of the patients presented with liver or brain subtypes, and the proportion of male patients reached 62.03%; specifically, it includes structured fields such as general information, blood routine, urine copper, serum ceruloplasmin, and Child-Pugh score-related indicators. The original data was cleaned, and outliers and invalid values ​​were removed.

[0183] (2) For missing values, imputation is performed using nearby valid samples; for numerical features, min-max normalization is used to map them to the [0,1] interval to eliminate the differences in the scale of different features. Taking Wilson's disease as an example, valid samples that meet the modeling conditions can be screened from the original case records for subsequent training and validation.

[0184] The second step involves using the mutual information feature selection method to calculate the mutual information score between each conditional feature and the diagnostic label, and then filtering the feature subsets according to the threshold.

[0185] This invention calculates a mutual information score MI(X,Y) between each conditional feature X and diagnostic label Y. Mutual information quantifies the statistical dependence strength between features and labels, simultaneously characterizing linear and nonlinear relationships. The mutual information scores of each feature are sorted, and a threshold is determined using the quartile method. Features with low contribution are removed, retaining key features with mutual information scores above the threshold, resulting in a feature subset. This step reduces redundancy and noise input, improving the efficiency and robustness of subsequent classification models. This invention further optimizes the initial 50 features. Through feature simplification based on mutual information, the number of features was ultimately reduced to 23 conditional features, including: Kaiser-Fleisher ring (KF ring), Child-Pugh score, urinary copper, red blood cell count (RBC), platelet count (PLT), prothrombin time (PT), international normalized ratio (INR), activated partial thromboplastin time (APTT), fasting blood glucose (FBG), alanine aminotransferase (ALT), aspartate aminotransferase (AST), lactate dehydrogenase (LDH), triglycerides (TG), total cholesterol (TC), low-density lipoprotein cholesterol (LDL), albumin (ALB), prealbumin (PAB), total bilirubin (TBIL), alkaline phosphatase (AKP), blood urea nitrogen (BUN), type IV collagen (CIV), hyaluronic acid (HA), laminin (LN), and one decision feature—Wilson's disease subtype.

[0186] Table 1 Skewness and Kurtosis of Conditional Attributes

[0187] Copper in urine 1.8133 7.3707 RBC -0.6369 1.0026 PLT 1.2319 2.2610 PT 6.4649 92.2960 INR 5.7095 76.7001 APTT 1.5680 4.8313 FBG 1.1402 3.7017 ALT 5.0532 42.5501 AST 6.4701 73.6537 LDH 2.0645 9.7878 TG 6.1306 93.1713 TC 1.5178 4.7101 LDL 2.0488 9.3817 ALB -1.8154 5.3459 PAB -0.5920 0.6516 TBIL 5.0346 32.5968 AKP 2.8589 9.5350 BUN 1.8469 22.0935 CIV 3.9036 30.3590 HA 4.4199 30.3221 LN 2.5496 9.0847

[0188] The third step is to perform class balancing on the training samples corresponding to the feature subset, and use the SMOTE algorithm to perform synthetic oversampling on the minority classes.

[0189] (1) Category statistics and determination of target sample size. The number of samples for each Wilson's disease subtype in the training dataset after mutual information filtering and normalization was counted, and the six subtypes were denoted as:

[0190] .

[0191] The corresponding sample size is ,make As the target sample size; for each minority class The number of samples to be synthesized is .

[0192] (2) Feature space and distance metric. Let each preprocessed patient sample be represented as a d-dimensional feature vector: Each component represents a clinical indicator that has undergone missing value imputation, outlier handling, and min-max normalization. The Euclidean distance between samples is calculated using the following formula:

[0193] .

[0194] (3) For each minority class The original sample In the same category Within the sample set, calculate its Euclidean distance according to the above-mentioned Euclidean distance metric. The nearest neighbor sample set:

[0195] ,

[0196] in For example, the preset number of nearest neighbors. .

[0197] (4) For each minority class sample When a new synthetic sample needs to be generated, it is derived from its nearest neighbor set. Randomly select a neighboring sample A random interpolation coefficient λ, taking values ​​in the interval [0,1], is generated, and a new synthetic sample is generated according to the following linear interpolation formula. :

[0198] ,

[0199] That is, for each feature m, we have:

[0200] ,

[0201] The above generation process is carried out in the normalized feature space to ensure that each dimension of the new sample still falls within the preset value range.

[0202] (5) For each minority class c, repeat steps (53-54) to generate synthetic samples for all its original samples in a loop until the cumulative number of synthetic samples for that class reaches the required number, so that the updated number of class samples satisfies: This allows the number of samples from the six Wilson's disease subtypes to reach an approximately balanced distribution in the training set.

[0203] (6) Merge the synthetic samples of all categories with the original samples to obtain a balanced training sample set:

[0204] ,

[0205] in Indicate category The corresponding set of all synthetic samples. The balanced training set was used for training multiple heterogeneous base classifiers (LightGBM, GBDT, Random Forest, SVM, XGBoost) to improve the recall rate of a few subtypes and the overall classification performance.

[0206] The fourth step is to train multiple heterogeneous base classifiers based on the balanced training samples to obtain each base classification model.

[0207] This invention trains five heterogeneous classifiers: LightGBM, GBDT, Random Forest, SVM, and XGBoost.

[0208] 1. LightGBM is configured with 31 leaf nodes, 100 trees, a learning rate of 0.1, and a maximum depth of 3.

[0209] 2. GBDT learning rate 0.1, weak learner 160;

[0210] 3. Randomize the forest tree count to 150 and use Gini impurity;

[0211] 4. The SVM uses an RBF kernel, a regularization parameter of 1.0, and enables probabilistic prediction;

[0212] 5. Maximum depth of XGBoost; 6. Learning rate of 0.3 and use of log loss.

[0213] The performance of each classifier is as follows Figure 2-4 As shown.

[0214] (1) The training sample set after data preprocessing, mutual information feature filtering, and SMOTE class balancing is denoted as: ,

[0215] in:

[0216] Indicates the first Feature vectors of a patient sample;

[0217] This indicates the number of feature dimensions retained after filtering;

[0218] This indicates the Wilson's disease subtype label corresponding to the patient; the subtype label includes six categories: hepatic, neurological, asymptomatic, arthritic, cutaneous, and renal.

[0219] During the training phase, the balanced sample set is divided into a training set and a validation / test set for model fitting and performance evaluation.

[0220] (2) In order to improve the generalization ability and stability of the model in multi-class auxiliary diagnosis scenarios, multiple classification algorithms with different structures and learning mechanisms were selected as heterogeneous base classifiers to form a set of base classifiers:

[0221] ,

[0222] These correspond to the following classifiers: LightGBM classifier, GBDT classifier, Random Forest classifier, Support Vector Machine (SVM) classifier, and XGBoost classifier, respectively.

[0223] Heterogeneous models are used because different classifiers have different advantages in nonlinear relationships, feature interactions, multi-class boundaries, minority class recognition, and robustness. They can be integrated after weights are calculated using TOPSIS.

[0224] LightGBM is an ensemble classifier based on gradient boosting decision trees. It employs a histogram-based splitting strategy, discretizing continuous features into multiple buckets to reduce the number of candidate split points, thereby lowering training complexity and improving training efficiency. Its specific training process is as follows:

[0225] 1-1) Initialize the model output;

[0226] 1-2) Calculate the gradient of the loss function of the current model on the training samples;

[0227] 1-3) Train a new decision tree using the negative gradient direction as the fitting target;

[0228] 1-4) Add the decision tree to the existing tree set;

[0229] 1-5) Update the overall model output;

[0230] 1-6) Repeat the above process until the preset number of trees or the convergence condition is reached.

[0231] If the first If the wheel model is denoted as , then its recursive form can be expressed as:

[0232] ,

[0233] in: For the first A new tree was added. This is the learning rate.

[0234] GBDT is a gradient boosting decision tree model, and its specific training steps include:

[0235] 2-1) Initialize the classification model with a constant function;

[0236] 2-2) For each iteration, calculate the residual or negative gradient of the current model on each training sample;

[0237] 2-3) Train a new regression tree / classification tree based on the residuals;

[0238] 2-4) Multiply the new tree by the learning rate and then add it to the current model;

[0239] 2-5) Output the final classifier after reaching the preset number of iterations.

[0240] Its recursive expression can also be expressed as: ,

[0241] in: For the first Wheel integration model, To create a new tree for fitting the residuals, This is the learning rate.

[0242] Random forests belong to the Bagging ensemble model category. They construct multiple independent decision trees in parallel and merge the results of each tree through voting to reduce the variance problem caused by overfitting of a single tree. The specific training process includes:

[0243] 3-1) Generate multiple Bootstrap sample subsets from the training sample set using sampling with replacement;

[0244] 3-2) Train a decision tree for each Bootstrap sample subset;

[0245] 3-3) When splitting at each node, the optimal splitting variable is selected only from a randomly selected subset of features;

[0246] 3-4) Repeat the above process to generate a predetermined number of decision trees;

[0247] 3-5) Perform majority voting on the output of all trees to obtain the final predicted category.

[0248] If the first The output of the tree is Then the overall output of the random forest can be expressed as:

[0249] ,

[0250] in: The number of decision trees; This indicates the majority vote.

[0251] The specific training steps for a Support Vector Machine (SVM) classifier are as follows:

[0252] 4-1) Map the input feature vectors to the kernel space;

[0253] 4-2) Construct the maximum margin optimization problem;

[0254] 4-3) Solve for the support vectors and their Lagrange multipliers;

[0255] 4-4) Obtain the optimal classification hyperplane;

[0256] 4-5) For multi-class classification tasks, use a one-to-many or one-to-one strategy to output the classification results for each class;

[0257] 4-6) After enabling probability output, the predicted probabilities of each category can be obtained.

[0258] The optimization objective of SVM can be written as:

[0259] ,

[0260] Satisfy constraints:

[0261] ,

[0262] in: For kernel mapping; For regularization parameters; These are slack variables.

[0263] XGBoost is an improved gradient boosting tree model that introduces a regularization term into the traditional boosting tree model and supports parallel computation, tree pruning, and efficient feature splitting search. Its training process includes:

[0264] 5-1) Initialize the model predictions;

[0265] 5-2) Calculate the first and second gradients based on the current prediction results;

[0266] 5-3) Select the optimal split point based on the objective function gain;

[0267] 5-4) Generate new trees and prune them;

[0268] 5-5) Add the new tree to the existing model;

[0269] 5-6) Repeat the iterations until the maximum number of trees is reached or the convergence condition is met. The overall objective function can be expressed as:

[0270] ,

[0271] in: The loss function; For the first Tree; This is a regularization term used to control model complexity.

[0272] (3) After the above five basic classifiers are trained, corresponding basic classification models are formed respectively:

[0273] ,

[0274] For any input patient sample X, each classification model outputs a predicted class, or a probability distribution about the six WD subtypes: ,

[0275] Where: k represents the k-th base classifier; L represents the total number of classes; and represents the probability that the classifier predicts a sample belongs to the j-th class. Each base classifier selects the class with the highest probability as its single-model output:

[0276] .

[0277] This output will be used in subsequent steps for performance evaluation, decision matrix construction, TOPSIS weight calculation, and weighted voting fusion.

[0278] The fifth step is to calculate multiple classification metrics for each basic classification model on the validation or test set, and construct a decision matrix with the classification metrics as attributes.

[0279] (1) The multiple heterogeneous basic classification models trained in step four are denoted as follows: In this embodiment: The classifiers used are LightGBM, GBDT, Random Forest, SVM, and XGBoost.

[0280] The sample set to be evaluated is denoted as: ,

[0281] Where: is the first Feature vectors of each test sample; This corresponds to the actual category label; the category label is one of the six subtypes of Wilson's disease.

[0282] For any basic classification model Test samples After inputting the data into the model, we obtain its prediction results:

[0283] ,

[0284] Or obtain the corresponding category probability output: ,in: The total number of categories, Indicates the first The classification model predicts the first... The sample belongs to the first The probability of each class. After each classification model completes its predictions for all test samples, it generates a corresponding set of prediction results, which are used for subsequent performance evaluation.

[0285] (2) For each basic classification model, construct the confusion matrix corresponding to the model based on the true label and the predicted label:

[0286] ,

[0287] in: Indicates the first In the nth model, the true category is the nth Class and predicted as the first The number of samples in each class; This represents the dimensions of the multi-class confusion matrix.

[0288] (3) After obtaining the confusion matrix of each model, six performance evaluation metrics are calculated as attribute columns of the subsequent decision matrix. The six metrics used are: Accuracy, Precision, Recall, F1Score, Specificity, and AUROC (Area Under the ROC Curve). Their calculation formulas are as follows:

[0289] Accuracy represents the proportion of correctly classified samples out of the total number of test samples. It can be expressed as:

[0290] ,

[0291] in: Indicates the first The number of correctly classified samples of a class; This indicates the total number of test samples.

[0292] Precision, for the first Class, precision is defined as:

[0293] ,

[0294] in: This indicates that it was incorrectly predicted as the first. The number of samples in each class. For multi-class classification problems, a weighted average is used to obtain the overall precision.

[0295] Recall, for the first Class, recall rate is defined as:

[0296] ,

[0297] in: Indicates that it truly belongs to the first The number of samples that were classified as one class but were predicted as other classes. The overall recall rate for multi-class classification can be calculated using a weighted average.

[0298] The F1 score is used to comprehensively reflect the balance between precision and recall. For the first... kind: The overall F1 score can be obtained by weighted averaging.

[0299] Specificity measures a model's ability to identify negative class samples. For the first... A class can be defined as:

[0300] ,

[0301] in: Indicates that it does not belong to the first The number of samples that are correctly identified as not belonging to a class. In multi-class scenarios, the overall specificity can be the average of the specificities of each class.

[0302] AUROC (Area Under the ROC Curve) measures a model's ability to distinguish between different classes. In multi-class classification problems, a one-to-many strategy is used to calculate the area under the ROC curve for each class separately, and then the average is taken as the overall AUROC.

[0303] (4) For each basic classification model The above six performance metrics on the test set are combined into a performance evaluation vector: ,in: Indicates accuracy; Indicates accuracy; Indicates recall rate; Indicates the F1 score; Indicates specificity; This represents AUROC. In other words, each base classifier is ultimately mapped to a six-dimensional performance description vector.

[0304] (5) All The performance evaluation vectors of the base classifiers are arranged in rows to construct the original decision matrix:

[0305] .

[0306] In this implementation: number of rows There are five basic classifiers; the number of columns is 6, corresponding to six performance metrics.

[0307] (6) In order to reduce the impact of accidental division or fluctuations in a single experiment, in one embodiment, each basic classification model is repeatedly trained and tested. In each experiment, six evaluation indicators are recorded, and the mean and standard deviation of each indicator are calculated to form a more stable evaluation result.

[0308] If the first The model of the first The indicator in the first The values ​​taken in this experiment are denoted as follows: Then its mean and standard deviation are respectively:

[0309]

[0310] ,

[0311] in: To repeat the experiment a certain number of times, further intervals were used. This indicates the range of fluctuation of the model on this evaluation index, thus reflecting the stability of the model. The decision matrices and proximity of each base classifier are shown in Table 2.

[0312] Table 2 Decision matrices and proximity scores of each base classifier

[0313] LightGBM [0.9046,0.9247] [0.9050,0.9246] [0.9046,0.9247] [0.9044,0.9244] [0.9816,0.9854] [0.9872,0.9907] GBDT [0.8438,0.8676] [0.8421,0.8668] [0.8438,0.8676] [0.8423,0.8668] [0.9704,0.9748] [0.9716,0.9780] RF [0.9025,0.9237] [0.9028,0.9237] [0.9025,0.9237] [0.9023,0.9034] [0.9812,0.9852] [0.9869,0.9906] SVM [0.7209,0.7501] [0.7162,0.7485] [0.7209,0.7501] [0.7057,0.7373] [0.9477,0.9529] [0.9248,0.9394] XGBoost [0.9037,0.9236] [0.9030,0.9230] [0.9034,0.9232] [0.9029,0.9228] [0.9813,0.9851] [0.9866,0.9908] WVC [0.9192,0.9316] [0.9189,0.9314] [0.9192,0.9316] [0.9187,0.9312] [0.9842,0.9867] [0.9896,0.9918]

[0314] The sixth step is to use the TOPSIS method to calculate the voting weights of each basic classification model.

[0315] (1) Assume there is a total There are 1 basic classification model, and the number of evaluation indicators is 1. The original decision matrix is ​​constructed as follows: ,in , indicating the first A basic classification model; , , indicating the first One evaluation indicator.

[0316] (2) Since different evaluation indicators have different dimensions, the decision matrix needs to be dimensionless. Vector normalization (a standard practice in TOPSIS) is used:

[0317] ,

[0318] Obtain the standardized matrix: .

[0319] (3) Setting the weights of the indicators: In the corresponding embodiments of the present invention, the importance of each evaluation indicator to the model performance can be set to equal weight, or the weight vector can be set according to empirical knowledge: ,satisfy The weighted normalized matrix is ​​obtained by weighting the normalized matrix: , denoted as: .

[0320] (4) Determine the positive and negative ideal solutions. Since all evaluation indicators are benefit-type indicators (the larger the value, the better), therefore:

[0321] Positive ideal solution: ,in: This represents the combined vector when each indicator achieves its optimal performance.

[0322] Negative ideal solution: ,in: This represents the combined vector when each indicator reaches its worst performance.

[0323] (5) For each basic classification model Calculate the Euclidean distance between it and the positive ideal solution and the negative ideal solution.

[0324] Distance from the ideal solution: ,

[0325] Distance to negative ideal solution: .

[0326] (6) Calculate the relative proximity coefficient. According to the TOPSIS principle, the first... The relative proximity of the base classifiers is:

[0327] ,

[0328] in: , The larger the value, the closer the classifier is to the ideal solution.

[0329] (7) Compose a weight vector from the relative proximity coefficients of all basic classification models:

[0330] ,

[0331] And it can be further normalized: This yields the classifier weights used for the final weighted voting.

[0332] The seventh step involves obtaining the voting weights of each basic classification model, and then performing a weighted voting fusion based on the weights of the patient samples to be diagnosed, thereby obtaining the final Wilson's disease diagnosis category or subtype result.

[0333] (1) Feature construction of the patient to be diagnosed. Acquire the clinical multidimensional indicator data of the patients to be diagnosed and process it according to the same data processing procedure as in the training phase, including: missing value imputation, outlier or invalid value handling, min-max normalization, and mutual information filtering to obtain a feature subset. Let the feature vector of the patient to be diagnosed be represented as: ,in: These are the filtered feature dimensions.

[0334] (2) Basic classification model prediction output. Input the samples to be diagnosed into the pre-trained... Basic classification models: The prediction results of each model are obtained:

[0335] ,

[0336] in: In this embodiment, the possible categories or subtypes of Wilson's disease are indicated as follows: These correspond to the liver type, neurological type, asymptomatic type, joint type, skin type, and kidney type, respectively.

[0337] Each basic classification model can also output the probability distribution of belonging to each category:

[0338] ,

[0339] in: ,

[0340] The predicted category of the model can be determined by the maximum probability: .

[0341] (3) In step eight, the relative proximity coefficients of each basic classification model are calculated using the TOPSIS method: And form the model weight vector: ,in: Or, to further normalize it as: .

[0342] (4) Construct a category-weighted scoring function for each possible diagnostic category. Calculate its weighted voting score: ,

[0343] in: For indicator functions; when the first The model predicts the following categories: hour, ,otherwise: ,therefore: This represents all basic classification models for categories. The weighted support level.

[0344] (5) The final diagnostic category is obtained by selecting the category with the highest weighted score: ,in: This determines the final diagnostic category or subtype.

[0345] (6) Output diagnostic results. The final system output includes: Wilson disease diagnosis category or subtype, voting results of each model, model weights, and category weighted scores.

[0346] To clarify the independent contributions of each base learner in the ensemble framework, we conducted ablation experiments. Using a full ensemble model comprising SVM, GBDT, Random Forest (RF), LightGBM, and XGBoost as a baseline, we constructed ablation models by progressively removing one or more learners. The experiments were evaluated using six metrics: accuracy, recall, specificity, precision, F1 score, and AUROC. As shown in Table 3, the full ensemble model consistently outperformed all scaled-down comparison models, fully demonstrating the significant advantages of multi-algorithm ensembles in improving model stability and generalization ability.

[0347] It is noteworthy that the combination of GBDT and SVM maintains strong predictive performance even without the involvement of other learners, indicating their high individual efficiency. Conversely, when key algorithms such as LightGBM and XGBoost are excluded, classification performance drops significantly, with some combinations achieving an accuracy as low as 0.8962 and an F1 score as low as 0.9799, highlighting their crucial role in ensemble models.

[0348] Table 3 Evaluation results of different models

[0349] Full Model 0.9264 0.9262 0.9264 0.9260 0.9856 0.9898 SGRX 0.9204 0.9204 0.9204 0.9198 0.9845 0.9892 SRLX 0.9234 0.9236 0.9234 0.9230 0.9851 0.9906 SGLX 0.9214 0.9210 0.9214 0.9210 0.9847 0.9887 GRLX 0.9224 0.9221 0.9224 0.9219 0.9849 0.9897 SGRL 0.9204 0.9201 0.9204 0.9200 0.9845 0.9890 SRX 0.9234 0.9233 0.9234 0.9230 0.9851 0.9903 SGX 0.9163 0.9164 0.9163 0.9159 0.9837 0.9874 GRX 0.9194 0.9190 0.9194 0.9188 0.9843 0.9897 SGR 0.8962 0.8957 0.8962 0.8955 0.9799 0.9858 SLX 0.9234 0.9235 0.9234 0.9230 0.9851 0.9895 RLX 0.9254 0.9250 0.9254 0.9250 0.9854 0.9901 SRL 0.9224 0.9230 0.9224 0.9223 0.9849 0.9898 GLX 0.9224 0.9220 0.9224 0.9219 0.9849 0.9888 SGL 0.9133 0.9131 0.9133 0.9130 0.9831 0.9870 GRL 0.9204 0.9200 0.9204 0.9200 0.9845 0.9890 SX 0.9183 0.9187 0.9183 0.9180 0.9841 0.9889 RX 0.9254 0.9253 0.9254 0.9251 0.9855 0.9905 SR 0.9214 0.9222 0.9214 0.9215 0.9847 0.9878 LX 0.9234 0.9235 0.9234 0.9230 0.9851 0.9898 SL 0.9204 0.9203 0.9204 0.9200 0.9845 0.9881 GL 0.9163 0.9158 0.9163 0.9159 0.9837 0.9870

[0350] To comprehensively evaluate the effectiveness of the proposed ensemble prediction model, we designed and conducted comparative experiments. Representative single-machine learning models such as Decision Tree (DT), Extreme Learning Machine (ELM), Logistic Regression (LR), and K-Nearest Neighbors (KNN) were selected as benchmarks. All models were trained and tested on the same dataset and used the exact same preprocessing procedures to ensure fairness and comparability of the results. Furthermore, the six mainstream classification metrics listed in Table 3 were used to evaluate model performance from multiple dimensions; the results are detailed in Table 4.

[0351] Table 4 Comparison results with existing methods

[0352] Decision Tree 0.8044 0.8135 0.9605 0.8098 0.8107 0.8870 KNN 0.8286 0.8392 0.9656 0.8342 0.8150 0.9614 Logistic Regression 0.7006 0.7125 0.9402 0.6908 0.6928 0.8861 ELM 0.7389 0.7529 0.9476 0.7240 0.7302 0.9221 Full Model 0.9264 0.9262 0.9264 0.9260 0.9856 0.9898

[0353] In addition, in the Kappa coefficient test, we evaluated the classification consistency between five basic classifiers (LightGBM, GBDT, RF, SVM, XGBoost) and the ensemble learning model. The results are as follows: Figure 5 As shown, the ensemble learning model has the highest Kappa coefficient, indicating that it can effectively integrate the advantages of various basic classifiers, thereby improving the overall classification performance.

[0354] Overall evaluation results show that, compared with traditional algorithms, the developed model performs better in terms of accuracy, stability, AUC, and other indicators, and can provide doctors with accurate basis for diagnosing subtypes of Wilson's disease.

[0355] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A rare disease auxiliary diagnostic method based on EL-TOPSIS weighted decision-making, characterized in that, Includes the following steps: Step 1: Obtain clinical multidimensional indicator data of patients with the target rare disease and construct a training dataset containing sample data of multiple patients with the target rare disease; Step 2: Perform data preprocessing on the training dataset described in Step 1. The preprocessing methods include at least missing value imputation, outlier / invalid value handling, and numerical feature normalization. Step 3: For the training dataset after data preprocessing in Step 2, the mutual information feature selection method is used to calculate the mutual information score between each conditional feature and the diagnostic label, and the feature subset is obtained by filtering according to the threshold. Step four: Perform class balancing on the training samples corresponding to the feature subset described in step three, and use the SMOTE algorithm to perform synthetic oversampling on the minority classes; Step 5: Train multiple heterogeneous base classifiers based on the training samples balanced in Step 4 to obtain each base classification model; Step 6: Calculate multiple classification metrics for each basic classification model on the validation set or test set, and construct a decision matrix with the classification metrics as attributes. Step 7: Standardize and weight the decision matrix described in Step 6 using the TOPSIS method, determine the positive and negative ideal solutions, and calculate the relative proximity coefficient. The relative proximity coefficient is then used as the voting weight of the corresponding basic classification model. Step 8: Feature construction of the samples to be diagnosed. The samples to be diagnosed are processed in the same way as the data processing in the training phase. The prediction results are then output by each basic classification model. Weighted voting is performed based on the voting weights to output the diagnosis category or subtype of the rare disease.

2. The rare disease auxiliary diagnosis method based on EL-TOPSIS weighted decision-making according to claim 1, characterized in that: Step one, the clinical multidimensional indicator data, includes multidimensional clinical data of patients obtained from electronic medical record systems or clinical databases, specifically including but not limited to the following types of data: General information: including the patient's gender, age, medical history, family history, and past medical records; Laboratory testing indicators: liver function tests, urinary copper content, serum ceruloplasmin levels, and other medical chemical indicators; Physical signs scoring indicators: including the patient's clinical signs and the physician's clinical assessment; Structured clinical examination results: Imaging examination results, genetic testing results, MRI and CT imaging data, which are converted into numerical or categorical forms. These data will be entered in the form of structured data tables and will be appropriately processed and transformed according to the type of data. The diagnostic label is the disease category or subtype of the target rare disease.

3. The rare disease auxiliary diagnosis method based on EL-TOPSIS weighted decision-making according to claim 1, characterized in that: The preprocessing method in step two is as follows: Missing value imputation: For missing data, imputation is performed using valid data from neighboring samples; Outlier and invalid value removal: By setting a threshold, obvious outliers are removed; Min-Max Normalization: Standardize all numerical features by using a mapping method between minimum and maximum values ​​to normalize all feature values ​​to the [0,1] interval, thereby eliminating the dimensional differences between different features and ensuring that each feature has the same weight during training.

4. The rare disease auxiliary diagnosis method based on EL-TOPSIS weighted decision-making according to claim 1, characterized in that: Step three specifically includes the following steps: 31) Calculate the mutual information score between each feature and the target variable to reflect the correlation between the feature and the diagnostic label; 32) Based on the scores, a threshold is determined using the quartile method to remove features with low contribution. The remaining features will be retained as the feature set used when training the model.

5. A rare disease auxiliary diagnostic method based on EL-TOPSIS weighted decision-making according to claim 1, characterized in that: Step four describes the specific steps involved in class balancing using the SMOTE algorithm: 41) Category Statistics and Target Sample Size Determination: The sample size of each rare disease subtype in the training dataset after mutual information filtering and normalization is statistically analyzed, and the six subtypes are denoted as: ； The corresponding sample size is ,make As the target sample size; for each minority class The number of samples to be synthesized is ; 42) Feature space and distance metric: Let each preprocessed patient sample be represented as a d-dimensional feature vector. Each component represents a clinical indicator that has undergone missing value imputation, outlier handling, and min-max normalization. The distance between samples is calculated using the Euclidean distance formula: ； 43) For each minority class The original sample In the same category Within the sample set, calculate its Euclidean distance according to the above-mentioned Euclidean distance metric. The nearest neighbor sample set: ， in The preset number of nearest neighbors, ; 44) For each minority class sample When a new synthetic sample needs to be generated, it is derived from its nearest neighbor set. Randomly select a neighboring sample A random interpolation coefficient λ, taking values ​​in the interval [0,1], is generated, and a new synthetic sample is generated according to the following linear interpolation formula. : ， That is, for each feature m, we have: ， The above generation process is carried out in the normalized feature space to ensure that each dimension of the new sample still falls within the preset value range; 45) For each minority class c, repeat steps 43 and 44 to iteratively generate synthetic samples for all its original samples until the cumulative number of synthetic samples for that class reaches the required number, so that the updated number of class samples satisfies: This allows the number of samples from the six rare disease subtypes to achieve an approximately balanced distribution in the training set. 46) Merge the synthetic samples of all categories with the original samples to obtain a balanced training sample set: ， in Indicate category The corresponding set of all synthetic samples, after being balanced, is used for training multiple heterogeneous base classifiers, thereby improving the recall rate of a few subtypes and the overall classification performance.

6. A rare disease auxiliary diagnostic method based on EL-TOPSIS weighted decision-making according to claim 5, characterized in that: The process of training multiple heterogeneous base classifiers based on the balanced training samples to obtain base classification models including at least LightGBM, GBDT, Random Forest, SVM, and XGBoost, specifically includes the following steps: 1) The training sample set after data preprocessing, mutual information feature filtering, and SMOTE class balancing is denoted as: , in: Indicates the first Feature vectors of a patient sample This indicates the number of feature dimensions retained after filtering. This indicates the rare disease subtype label corresponding to the patient; the subtype labels include six categories: hepatic, neurological, asymptomatic, arthritic, cutaneous, and renal. During the training phase, the balanced sample set is divided into a training set and a validation / test set for model fitting and performance evaluation. 2) Select multiple classification algorithms with different structures and learning mechanisms as heterogeneous base classifiers to form a base classifier set: ， These correspond to: ƒ1-LightGBM classifier, ƒ2-GBDT classifier, ƒ3-Random Forest classifier, ƒ4-Support Vector Machine (SVM) classifier, and ƒ5-XGBoost classifier, respectively. LightGBM is an ensemble classifier based on gradient boosting decision trees. It employs a histogram-based splitting strategy, discretizing continuous features into multiple buckets to reduce the number of candidate split points, thereby lowering training complexity and improving training efficiency. Its specific training process is as follows: 1-1. Initialize model output. 1-2. Calculate the gradient of the loss function of the current model on the training samples. 1-3. Using the negative gradient direction as the fitting target, train a new decision tree. 1-4. Add the decision tree to the existing tree set. 1-5. Update the overall model output. 1-6. Repeat the above process until the preset number of trees or the convergence condition is reached. If the wheel model is denoted as , then its recursive form is: ， in: For the first A new tree was added. The learning rate; GBDT is a gradient boosting decision tree model, and its specific training steps include: 2-1. Initialize the classification model with a constant function. 2-2. For each iteration, calculate the residual or negative gradient of the current model on each training sample. 2-3. Train a new regression tree / classification tree based on the residuals. 2-4. Multiply the new tree by the learning rate and then add it to the current model. 2-5. After reaching the preset number of iterations, output the final classifier, whose recursive expression is: ， in: For the first Wheel integration model, To create a new tree for fitting the residuals, The learning rate; Random forests belong to the Bagging ensemble model category. They construct multiple independent decision trees in parallel and merge the results of each tree through voting to reduce the variance problem caused by overfitting of a single tree. The specific training process includes: 3-1. Generate multiple Bootstrap sample subsets from the training sample set using sampling with replacement. 3-2. Train a decision tree for each Bootstrap sample subset. 3-3. When splitting at each node, the optimal splitting variable is selected only from a randomly selected subset of features. 3-4. Repeat the above process to generate a predetermined number of decision trees. 3-5. Perform majority voting on the output of all trees to obtain the final predicted category. If the first tree... The output of the tree is Then the overall output of the random forest is: ， in: The number of decision trees; Indicates the majority vote; The specific training steps for a Support Vector Machine (SVM) classifier are as follows: 4-1. Map the input feature vector to the kernel space. 4-2. Construct the maximum margin optimization problem. 4-3. Solve for the support vectors and their Lagrange multipliers. 4-4. Obtain the optimal classification hyperplane. 4-5. For multi-class classification tasks, use a one-to-many or one-to-one strategy to output the classification results for each class. 4-6. After enabling probability output, the prediction probability for each category can be further obtained. The SVM optimization objective is: ， Satisfy constraints: ， in: For kernel mapping; For regularization parameters; These are slack variables; XGBoost is an improved gradient boosting tree model that introduces a regularization term into the traditional boosting tree model and supports parallel computation, tree pruning, and efficient feature splitting search. Its training process includes: 5-1. Initialize model predictions. 5-2. Calculate the first and second gradients based on the current prediction results. 5-3. Select the optimal split point based on the objective function gain. 5-4. Generate new trees and prune them. 5-5. Add the new tree to the existing model. 5-6. Repeat the iteration until the maximum number of trees is reached or the convergence condition is met. The overall objective function is: ， in: The loss function; For the first Tree; This is a regularization term used to control model complexity; 3) After the above five basic classifiers are trained, corresponding basic classification models are formed: ， For any input patient sample X, each classification model outputs a predicted class, or a probability distribution for the six rare disease subtypes: ， Where: k represents the k-th base classifier; L represents the total number of classes; and represents the probability that the classifier predicts a sample belongs to the j-th class. Each base classifier selects the class with the highest probability as its single-model output. ， This output will be used in subsequent steps for performance evaluation, decision matrix construction, TOPSIS weight calculation, and weighted voting fusion.

7. A rare disease auxiliary diagnostic method based on EL-TOPSIS weighted decision-making according to claim 6, characterized in that: Step six involves calculating multiple classification metrics for each basic classification model on the validation or test set, and constructing a decision matrix with the classification metrics as attributes. This specifically includes the following steps: 1) The multiple heterogeneous basic classification models trained in step six are denoted as follows: Where K represents the number of heterogeneous basic classification models. The sample set to be evaluated is denoted as: , Where: is the first Feature vectors of each test sample; This corresponds to the actual category label; the category label is one of the six subtypes of rare diseases. For any basic classification model Test samples After inputting the data into the model, we obtain its prediction results: ， Or obtain the corresponding category probability output: ,in: The total number of categories, Indicates the first The classification model predicts the first... The sample belongs to the first The probability of a class; After each classification model completes its predictions for all test samples, it generates a corresponding set of prediction results for subsequent performance evaluation. 2) For each basic classification model, construct the confusion matrix corresponding to that model based on the true label and the predicted label: , in: Indicates the first In the nth model, the true category is the nth Class and predicted as the first The number of samples in each class; For the dimensions of the multi-class confusion matrix; 3) After obtaining the confusion matrix of each model, calculate six performance evaluation metrics as attribute columns of the subsequent decision matrix; the six metrics used are: Accuracy, Precision, Recall, F1Score, Specificity, and AUROC (area under the ROC curve). The calculation formulas are as follows: Accuracy represents the proportion of correctly classified samples out of the total number of test samples, expressed as: , in: Indicates the first The number of correctly classified samples of a class; Indicates the total number of test samples; Precision, for the first Class, precision is defined as: , in: This indicates that it was incorrectly predicted as the first. The number of samples in each class; for multi-class classification problems, a weighted average method is used to obtain the overall precision; Recall, for the first Class, recall rate is defined as: ， in: Indicates that it truly belongs to the first The total recall of a multi-class system can be calculated using a weighted average method, which measures the number of samples that are classified as one class but are predicted as other classes. The F1 score is used to comprehensively reflect the balance between precision and recall. kind: The overall F1 score can be obtained by weighted averaging. Specificity measures a model's ability to identify negative class samples. The class is defined as: , in: Indicates that it does not belong to the first The number of samples that are correctly identified as not belonging to a class; in multi-class scenarios, the overall specificity can be the average of the specificities of each class. AUROC (Area Under the ROC Curve) is used to measure the ability of a model to distinguish between different classes. In multi-class classification problems, a one-to-many strategy is used to calculate the area under the ROC curve for each class separately, and then the average value is taken as the overall AUROC. 4) For each basic classification model The above six performance metrics on the test set are combined into a performance evaluation vector: ,in: Indicates accuracy; Indicates accuracy; Indicates recall rate; Indicates the F1 score; Indicates specificity; Indicates AUROC; 5) All The performance evaluation vectors of the base classifiers are arranged in rows to construct the original decision matrix: ， In this implementation: number of rows There are five basic classifiers; the number of columns is 6, corresponding to six performance metrics. 6) To reduce the impact of accidental classifications or fluctuations in a single experiment, each basic classification model was repeatedly trained and tested. Six evaluation indicators were recorded in each experiment, and the mean and standard deviation of each indicator were calculated to form more stable evaluation results. The model of the first The indicator in the first The values ​​taken in this experiment are denoted as follows: Then its mean and standard deviation are respectively: ， ， in: To repeat the experiment a certain number of times, further intervals were used. This indicates the range of fluctuation of the model on this evaluation index, thus reflecting the stability of the model.

8. A rare disease auxiliary diagnostic method based on EL-TOPSIS weighted decision-making according to claim 7, characterized in that: After constructing the performance decision matrix of the base classifiers, the TOPSIS method is used to calculate the voting weights of each base classification model. This process includes the following steps: 1) Assume there is a total There are 1 basic classification model, and the number of evaluation indicators is 1. The original decision matrix is ​​constructed as follows: ,in , indicating the first A basic classification model; , , indicating the first One evaluation indicator; 2) Since different evaluation indicators have different dimensions, the decision matrix needs to be dimensionless, using vector normalization: ， The standardized matrix is ​​obtained as follows: ; 3) Setting indicator weights: The importance of each evaluation indicator to model performance can be set to equal weight, or a weight vector can be set: ,satisfy The weighted normalized matrix is ​​obtained by weighting the normalized matrix: , denoted as: ; 4) Determine the positive and negative ideal solutions. Since all evaluation indicators are benefit-type indicators, as follows: Positive ideal solution: ,in: This represents the combined vector when each indicator achieves its optimal performance. Negative ideal solution: ,in: This represents the combined vector when each indicator reaches its worst performance. 5) For each basic classification model Calculate the Euclidean distance between it and the positive and negative ideal solutions. Distance from the ideal solution: , Distance to negative ideal solution: ; 6) Calculate the relative proximity coefficient. According to the TOPSIS principle, define the first... The relative proximity of the base classifiers is: ， in: , The larger the value, the closer the classifier is to the ideal solution; 7) Compile the relative proximity coefficients of all basic classification models into a weight vector: ， And it can be further normalized: This yields the classifier weights used for the final weighted voting.

9. A rare disease auxiliary diagnostic method based on EL-TOPSIS weighted decision-making according to claim 8, characterized in that: After obtaining the voting weights of each basic classification model, a weighted voting fusion based on weights is performed on the patient samples to obtain the final rare disease diagnosis category or subtype result. The specific steps include the following: 1) Construction of features for patients awaiting diagnosis: Obtain multidimensional clinical indicator data of patients awaiting diagnosis, and process them according to the same data processing procedures as in the training phase, including: imputation of missing values, handling of outliers or invalid values, min-max normalization, and mutual information filtering to obtain feature subsets. Let the feature vector of the patients awaiting diagnosis be represented as: ,in: These are the filtered feature dimensions; 2) Basic classification model prediction output. Input the samples to be diagnosed into the pre-trained... Basic classification models: The prediction results of each model are obtained: ， in: In this embodiment, the category or subtype of a rare disease is indicated as follows: These correspond to the liver type, neurological type, asymptomatic type, joint type, skin type, and kidney type, respectively. Each basic classification model can also output the probability distribution of belonging to each category: ， in: , The predicted category of the model can be determined by the maximum probability: ; 3) In step eight, the relative proximity coefficients of each basic classification model are calculated using the TOPSIS method: And form the model weight vector: ,in: Or further normalize to: ; 4) Construct a category-weighted scoring function for each possible diagnostic category. Calculate its weighted voting score: , in: For indicator functions; when the first The model predicts the following categories: hour, ,otherwise: ,therefore: This represents all basic classification models for categories. The weighted degree of support; 5) The final diagnostic category is obtained by selecting the category with the highest weighted score: ,in: The final output is the diagnostic category or subtype; 6) Output diagnostic results. The final system output results include: rare disease diagnosis category or subtype, voting results of each model, model weights, and category weighted scores.

Images