A system for predicting severity of coronary artery lesions in kawasaki disease

Abstract

The application discloses a system for predicting the severity of coronary artery lesions in Kawasaki disease, and relates to the technical field of medical data processing and machine learning. The system comprises: a data acquisition and preprocessing module for obtaining clinical raw data and generating a training sample set; a feature screening module that uses multiple feature screening algorithms of different principles to jointly screen and take the intersection to obtain a core prediction feature set; a stacking ensemble learning prediction module that performs multiple binary classification predictions through a first layer of base learners, and a second layer of meta learners that fuse the prediction probabilities into a four-level severity classification result; an explainability analysis module for calculating the contribution value of each feature to the prediction result and generating visual explanation information; and a Web application deployment module for packaging the system as a lightweight Web application. The application realizes four-level hierarchical prediction of the severity of coronary artery lesions, and improves the model explainability and clinical usability.

Description

Technical Field

[0001] This invention relates to the fields of medical data processing and machine learning technology, and in particular to a predictive system for the severity of coronary artery lesions in Kawasaki disease. Background Technology

[0002] Kawasaki disease (KD) is an acute, self-limiting vasculitis syndrome that primarily affects children. While its exact cause is not fully understood, it is generally believed to be closely related to abnormal activation of the immune system induced by infection. The most serious sequela of this disease is coronary artery lesions (CAL), including coronary artery dilation, aneurysm formation, and stenosis, which can lead to myocardial ischemia, myocardial infarction, and even sudden death, posing a serious threat to the long-term cardiovascular health of affected children. Epidemiological data show that approximately 20%–25% of untreated children with KD develop coronary artery damage, and even with standard intravenous immunoglobulin (IVIG) treatment, some children still retain varying degrees of coronary artery lesions. The American Heart Association (AHA) guidelines for the management of Kawasaki disease explicitly recommend risk stratification management for children with coronary artery lesions (CALs) based on severity: low-risk children require only short-term aspirin therapy and extended follow-up; intermediate-risk children require enhanced anticoagulation therapy and regular ultrasound follow-ups; and high-risk children require lifelong anticoagulation, intensive monitoring, and consideration of interventional procedures. Therefore, early assessment of the severity of coronary artery disease during the acute phase is of significant clinical importance for guiding individualized treatment, optimizing the allocation of medical resources, and improving the long-term prognosis of children with CALs.

[0003] Currently, scholars both domestically and internationally have constructed various predictive models for CAL (collateral arterial disease) in Kawasaki disease based on clinical data and machine learning algorithms. For example, some studies have used single machine learning algorithms such as logistic regression, random forest, or XGBoost to establish binary classification predictive models based on the demographic characteristics, peripheral blood laboratory indicators, and clinical manifestations of affected children, in order to distinguish whether CAL has occurred. However, these models have the following significant shortcomings: First, their predictive function is limited, only able to output a binary result of "whether CAL has occurred," unable to distinguish between four severity levels: normal, mild, moderate, and severe, making it difficult to meet the actual needs of stratified clinical treatment. Second, their algorithmic structure is simplistic; a single model has limited ability to fit high-dimensional, nonlinear clinical data, and when faced with class imbalance due to the scarcity of moderate and severe cases, predictive performance is prone to bias, resulting in weak generalization ability. Third, their interpretability is poor; the internal decision-making logic of traditional machine learning models is difficult to understand intuitively, and it is impossible to quantify the direction and magnitude of the influence of each input feature on the prediction result. Clinicians find it difficult to understand the basis for the prediction, reducing the trust and acceptance of the model in clinical practice. Fourth, most studies are based on single-center, small-sample data, lacking multi-center external validation, and the feature selection methods are crude. The use of a single selection algorithm leads to feature redundancy and high model complexity, which is not conducive to clinical promotion.

[0004] Based on publicly available patent and non-patent literature, existing related technologies can be mainly divided into two categories. One category is biomarker-based detection schemes, such as those that assess CAL risk by detecting specific microRNAs, platelet-monocyte aggregates, or specific gene expression levels in peripheral blood. These schemes rely on complex experimental equipment such as real-time quantitative PCR and flow cytometry, resulting in long detection cycles and high costs, making immediate assessment in clinical settings difficult. The other category is prediction systems based on clinical data and machine learning, such as Chinese patent application CN120496817A. This system integrates clinical manifestations and genetic data and uses a single machine learning model for prediction. However, this scheme still does not solve the problem of stratified prediction of CAL severity, nor does it introduce a class imbalance handling mechanism. Furthermore, due to the involvement of gene testing, its clinical application threshold remains high.

[0005] In summary, there is currently a lack of a predictive system that can stratify and predict the severity of coronary artery lesions based on routine clinical indicators, has good interpretability, and is easy to use at the bedside. Summary of the Invention

[0006] To address these issues, this invention provides a system for predicting the severity of coronary artery lesions in Kawasaki disease, which solves the problems in existing technologies such as the inability to achieve four-level stratified prediction of coronary artery lesion severity, low accuracy of single-model prediction, poor interpretability, and high threshold for clinical application.

[0007] To address the aforementioned technical problems, embodiments of the present invention provide a system for predicting the severity of coronary artery lesions in Kawasaki disease, the system comprising: The data acquisition and preprocessing module is used to acquire raw clinical data of children with Kawasaki disease, and to generate a training sample set after processing the raw clinical data by imputing missing values, standardization, class coding and class imbalance. The feature selection module is used to perform joint feature selection on the training sample set using multiple feature selection algorithms based on different principles, and to obtain the core prediction feature set by taking the intersection of the selection results of each algorithm. The Stacking ensemble learning prediction module is used to take the core prediction feature set as input. The Stacking ensemble learning prediction module includes a first base learner layer and a second meta learner layer. The first base learner layer performs multiple binary classification predictions on the severity of coronary artery lesions and outputs the probability of each binary classification prediction. The second meta learner layer merges the probability of each binary classification prediction into a four-level severity classification prediction result and outputs it. The interpretability analysis module is used to receive the prediction results output by the Stacking ensemble learning prediction module, calculate the contribution value of each feature in the core prediction feature set to the prediction results, and generate visual explanation information. The Web application deployment module is used to encapsulate the core prediction feature set, the Stacking integrated learning prediction module, and the interpretable analysis module into a lightweight Web application, providing feature input interface, prediction result display interface, and interpretable visualization display interface.

[0008] Preferably, in the data acquisition and preprocessing module, the class imbalance processing specifically includes: using synthetic minority class oversampling technology to amplify samples only in the training set corresponding to the moderate and severe categories of coronary artery lesion severity, so that the number of samples in the moderate and severe categories tends to be balanced with the number of samples in the majority category; the internal validation set, external validation set, and actual clinical prediction cases are not amplified.

[0009] Preferably, the feature filtering module specifically includes: The candidate feature matrix construction unit is used to construct a candidate feature matrix containing demographic features, clinical manifestations, laboratory indicators, and derived inflammatory indicators, using the preprocessed training set as input. The first screening unit is used to perform feature screening on the candidate feature matrix using a feature selection algorithm based on random forest, and output a first candidate feature set. The second screening unit is used to perform feature screening on the candidate feature matrix using a feature selection algorithm based on regularized regression, and output a second candidate feature set. The third screening unit is used to perform feature screening on the candidate feature matrix using a feature selection algorithm based on statistical testing, and output a third candidate feature set. The intersection determination unit is used to take the intersection of the first candidate feature set, the second candidate feature set, and the third candidate feature set, and determine the features that are simultaneously selected by multiple algorithms as the core prediction features.

[0010] Preferably, the core predictive features include: IVIG resistance, sex, platelet count, serum albumin, platelet-lymphocyte ratio, total bilirubin, CALLY index, and C-reactive protein.

[0011] Preferably, the Stacking ensemble learning prediction module further includes a severity label construction unit, which is used to calculate the coronary artery Z-score based on the coronary ultrasound examination results of the child. The formula for calculating the coronary artery Z-score is as follows: ; in, This indicates the actual inner diameter of the coronary artery segment as measured by ultrasound. This represents the normal predicted mean inner diameter of this coronary artery segment under the same body surface area conditions. This represents the prediction standard deviation for the coronary artery segment under the same body surface area conditions. The severity label construction unit compares the Z-values ​​of all coronary artery segments and takes the maximum value as the judgment criterion. When the maximum Z-value is less than 2.5, it is judged as normal; when 2.5 ≤ maximum Z-value < 5, it is judged as mild coronary artery disease; when 5 ≤ maximum Z-value < 10, it is judged as moderate coronary artery disease; and when the maximum Z-value ≥ 10, it is judged as severe coronary artery disease, and a four-level severity label is generated accordingly.

[0012] Preferably, the Stacking ensemble learning prediction module further includes a task transformation unit, which transforms the four severity labels into four one-to-many binary classification tasks, the four one-to-many binary classification tasks including: Normal vs. abnormal tasks: normal category samples are labeled as positive, and mild, moderate and severe category samples are labeled as negative. For both light and heavy tasks, light category samples are labeled as positive, while normal, moderate, and heavy category samples are labeled as negative. For moderate vs. non-moderate tasks, moderate category samples are labeled as positive, while normal, mild, and severe category samples are labeled as negative. For heavy vs. light tasks, heavy category samples are labeled as positive, while normal, light, and moderate category samples are labeled as negative. The first base learner layer is trained and predicted for the four one-to-many binary classification tasks respectively.

[0013] Preferably, the first base learner layer includes multiple candidate base learners, which are selected from any of the following: AdaBoost, CatBoost, decision tree, ExtraTrees, GBDT, KNN, LightGBM, random forest, SVM, and XGBoost. Each candidate base learner uses the features in the core prediction feature set as input variables and the label corresponding to the binary classification task as output variables. The first base learner layer also includes a hyperparameter optimization unit. This unit employs a Bayesian optimization method, using the combined score of MCC, F1-score, and PR-AUC as the objective function to search for and optimize the hyperparameters of each candidate base learner. The formula for calculating the combined score is as follows: Overall score = 0.4 × MCC + 0.3 × F1-score + 0.3 × PR-AUC.

[0014] Preferably, the second meta-learner layer includes: The probability vector generation unit is used to input the core prediction features of each child into the trained base learners corresponding to the four one-to-many binary classification tasks to obtain the prediction probability of normal category, mild category, moderate category and severe category, and combine the four prediction probabilities into a four-dimensional probability vector as secondary features. The training data generation unit is used to generate meta-learner training data in the training set using K-fold cross-validation. In each fold, the first-layer base learner is trained using K-1 folds, and the remaining 1 fold is used to generate the predicted probability of the sample in that fold. After repeating K times, a four-dimensional probability vector of all training samples is obtained. The meta-learner training unit is used to train a four-class fusion model by using a logistic regression model as the second-layer meta-learner, the four-dimensional probability vector as the input feature, and the true severity label of the child's coronary artery lesions as the output label. The probability normalization output unit is used to output the fusion probability of each child belonging to one of the four categories: normal, mild, moderate, and severe, and the category with the highest probability is used as the final predicted severity.

[0015] Preferably, the interpretability analysis module is implemented based on the SHAP method and includes: The SHAP value calculation unit is used to input the trained Stacking ensemble model, training set background data, and core prediction features into the SHAP interpreter, and calculate the SHAP value of each feature for the prediction result in each sample. The expression for the SHAP value is: ; in, The model represents the sample of children. The predicted output value, This represents the baseline prediction of the model. Indicates the first The contribution value of each feature to the prediction result. Indicates the feature number; The global feature importance analysis unit is used to average the absolute values ​​of each feature in all samples of the training set or validation set to obtain the overall importance of each feature, and output the feature importance ranking in descending order of overall importance; The individualized interpretation unit is used to calculate the SHAP value of each core feature for a single child to be predicted and sort them according to the size of the contribution value. It outputs the main factors that drive the prediction of the child to the corresponding severity and the direction of their influence. The visualization output unit is used to generate feature importance bar charts, SHAP waterfall charts, and SHAP force charts.

[0016] Preferably, the web application deployment module is built on the Streamlit framework and includes: The input interface unit is used to provide input interfaces for each feature in the core prediction feature set. The input interface supports binarized selection input of IVIG resistance, binarized selection input of gender, and numerical input of platelet count, serum albumin, platelet-lymphocyte ratio, total bilirubin, CALLY index and C-reactive protein. Each numerical input box is set with a reasonable numerical range verification. The background prediction unit is used to embed the Stacking integrated learning prediction module and the interpretability analysis module into the web application backend. In response to the prediction trigger operation after the input is completed, it calls the Stacking integrated learning prediction module to perform prediction calculation and calls the interpretability analysis module to generate corresponding visual explanation information. The results display unit is used to simultaneously display the four-level severity prediction probability distribution, the final severity prediction category, the SHAP waterfall plot, and the feature contribution ranking. The results export unit supports the export of prediction results and interpretability analysis reports for inclusion in the patient's medical record.

[0017] As can be seen from the above technical solutions, this invention application has the following beneficial effects: First, this invention employs a feature selection module that uses multiple feature selection algorithms based on different principles to perform joint selection and take the intersection. Combined with the Stacking ensemble learning prediction module, it constructs a two-layer ensemble learning architecture. The four-level classification task is decomposed into four one-to-many binary classification tasks. After selecting the optimal base learners specific to each task, they are fused through meta-learners. This effectively solves the technical problem in the prior art that it can only predict whether coronary artery lesions have occurred but cannot distinguish between the four severity levels of normal, mild, moderate, and severe. It achieves accurate stratified prediction of the severity of coronary artery lesions, providing direct decision support for individualized stratified treatment in clinical practice.

[0018] Secondly, this invention uses the SMOTE oversampling technique in the training set to amplify moderate and severe minority class samples through the data acquisition and preprocessing module. Combined with the minority class-sensitive hyperparameter optimization objective and multi-level optimal base learner selection rules in the Stacking ensemble learning prediction module, it effectively solves the class imbalance problem caused by the scarcity of moderate and severe cases in Kawasaki disease clinical data, and significantly reduces the risk of missed diagnosis of moderate and severe coronary artery lesions. At the same time, the interpretability analysis module uses the SHAP method to calculate the contribution value of each feature to the prediction results and generate visual explanation information, which makes up for the shortcomings of traditional machine learning models in terms of the interpretability of internal decision logic. This allows clinicians to intuitively understand the causes of the prediction results and enhances the trust and acceptability of the model in clinical practice.

[0019] Third, this invention encapsulates the core predictive features, Stacking ensemble model, and interpretability analysis module into a lightweight web application based on the Streamrit framework and deploys it on a cloud server through a web application deployment module. Doctors only need to input routine clinical indicators through a browser to complete real-time risk assessment in the clinic or ward, without the need for complex installation and professional equipment, which greatly reduces the threshold for clinical application. At the same time, through multi-center data partitioning and external validation mechanisms, the generalization ability and predictive reliability of the model in different clinical scenarios are effectively guaranteed, solving the problems of insufficient clinical translation and limited adaptability of existing technologies. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Referring to the drawings will make the features and advantages of the present invention clearer. The drawings are illustrative and should not be construed as limiting the present invention in any way. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a block diagram of a system for predicting the severity of coronary artery lesions in Kawasaki disease, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the feature filtering module in an embodiment of the present invention; Figure 3 This is a schematic diagram of the Stacking integrated learning prediction module in an embodiment of the present invention; Figure 4 This is a schematic diagram of the interpretability analysis module in an embodiment of the present invention; Figure 5 This is a structural diagram of the Web application deployment module in an embodiment of the present invention.

[0021] Explanation of reference numerals in the accompanying drawings: 100, Data Acquisition and Preprocessing Module; 200, Feature Selection Module; 210, Candidate Feature Matrix Construction Unit; 220, First Selection Unit; 230, Second Selection Unit; 240, Third Selection Unit; 250, Intersection Determination Unit; 300, Stacking Ensemble Learning Prediction Module; 310, Severity Label Construction Unit; 320, Task Transformation Unit; 330, First Base Learner Layer; 340, Second Meta Learner Layer; 350, Ordered Severity Probability Correction Unit; 400, Interpretability Analysis Module; 410, SHAP Value Calculation Unit; 420, Global Feature Importance Analysis Unit; 430, Individualized Interpretation Unit; 440, Visualization Output Unit; 500, Web Application Deployment Module; 510, Input Interface Unit; 520, Background Prediction Unit; 530, Result Display Unit; 540, Result Export Unit. Detailed Implementation

[0022] 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 only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1: To address the shortcomings of existing technologies, such as the inability to achieve a four-level stratified prediction of coronary artery lesion severity, low accuracy of single-model predictions, poor interpretability, and high barriers to clinical application, this invention proposes a prediction system for the severity of Kawasaki disease coronary artery lesions. Figure 1 As shown, the system includes: a data acquisition and preprocessing module 100, a feature selection module 200, a Stacking ensemble learning prediction module 300, an interpretability analysis module 400, and a Web application deployment module 500.

[0024] In this embodiment, the data acquisition and preprocessing module 100 is used to acquire the raw clinical data of children with Kawasaki disease, and after processing the raw clinical data for missing value imputation, standardization, class coding and class imbalance, a training sample set is generated.

[0025] 1. Data Acquisition: The data acquisition and preprocessing module 100 collected clinical data from 2,696 children diagnosed with Kawasaki disease from four tertiary hospitals in China (Hospital A, Hospital B, Hospital C, and Hospital D) between September 2020 and September 2024. The data covered 38 raw parameters in four categories: demographic characteristics, clinical manifestations, laboratory indicators, and imaging results.

[0026] Specifically, demographic characteristics included age, sex, weight, and date of admission; clinical manifestations included rash, conjunctival congestion, oral mucosal changes, edema of the extremities, perianal changes, and cervical lymphadenopathy; laboratory indicators included C-reactive protein (CRP), white blood cell count, neutrophil percentage, lymphocyte count and percentage, erythrocyte sedimentation rate, hemoglobin, platelet count (PLT), serum albumin (ALB), aspartate aminotransferase, alanine aminotransferase, serum sodium, potassium, calcium, creatine kinase isoenzyme MB, total bilirubin (TBIL), and derived inflammatory markers including CRP-lymphocyte ratio (CLR), neutrophil-lymphocyte ratio (NLR), platelet-lymphocyte ratio (PLR), neutrophil-platelet ratio (NPR), platelet-albumin ratio, systemic immune inflammatory index (SII), and CALLY index; and imaging results included coronary ultrasound examination results.

[0027] 2. Data filtering:

[0028] The data acquisition and preprocessing module 100 strictly followed the preset inclusion and exclusion criteria to screen valid data. The inclusion criteria were children with complete or incomplete Kawasaki disease who met the American Heart Association (AHA) criteria. Exclusion criteria included: use of glucocorticoids, immunosuppressants, or blood products within one month prior to initial IVIG treatment; receiving non-standard IVIG regimens; prior IVIG treatment before hospitalization; presence of severe complications; Kawasaki disease relapse; absence of fever or confirmed infection; presence of immunodeficiency or chromosomal abnormalities; and early discharge during treatment. After the above screening, data from 2696 valid children were obtained.

[0029] 3. Data partitioning: The data acquisition and preprocessing module 100 randomly divides the patient data from Hospital A into a training set (2156 cases) and an internal validation set (540 cases) at an 8:2 ratio; the patient data from the remaining three hospitals are used as an external validation cohort (1670 cases) to verify the model's generalization ability. The training set is used for model training and feature selection, the internal validation set is used for preliminary evaluation of model performance and reference for hyperparameter tuning, and the external validation set is used for the final evaluation of the model's generalization ability.

[0030] 4. Data preprocessing: The data acquisition and preprocessing module 100 performs the following preprocessing operations on the acquired data: ① Missing Value Handling: Multiple imputation is used to fill in missing data. Multiple imputation involves constructing multiple imputation datasets, imputing each dataset separately, and then combining the imputation results from all datasets to obtain the final imputed value. In practice, the MICE (Multivariate Imputation by Chained Equations) algorithm can be used. For each variable with missing values, a regression model or random forest model is built using other variables as predictors. The missing values ​​are iteratively predicted, and the process is repeated multiple times, with the average value taken as the final imputed value.

[0031] ② Data Standardization: Continuous variables are normalized using StandardScaler or MinMaxScaler from the Scikit-learn library. StandardScaler converts the data into a standard normal distribution with a mean of 0 and a standard deviation of 1; MinMaxScaler linearly scales the data to the [0,1] interval. Standardization eliminates the influence between different units of measurement, making each feature comparable in the model.

[0032] ③Category Coding: In order to enable clinical data to be recognized and processed by machine learning models, the data acquisition and preprocessing module 100 first determines the type of categorical variables based on their clinical meaning, number of values, and whether there is a natural order among the values, and then adopts the corresponding coding method.

[0033] Specifically, variables containing only two mutually exclusive categories are defined as binary variables, and are converted using a 0 / 1 binary coding method. Binary variables include gender, IVIG resistance, rash, conjunctival congestion, oral mucosal changes, limb edema, perianal changes, and cervical lymphadenopathy, etc., and their values ​​are only "yes / no," "male / female," or "present / absent." For example, male is coded as 1, female as 0; IVIG resistance is coded as 1, IVIG sensitivity as 0.

[0034] Variables containing three or more categories, where there is no clear hierarchical order among the categories, are defined as unordered categorical variables that are not binary classifications, and are transformed using one-hot encoding. For example, if the variable is a multi-category variable such as hospital origin, admission season, or other variable without a hierarchical relationship, multiple mutually exclusive 0 / 1 dummy variables are generated to prevent the model from misidentifying unordered categories as having a numerical relationship.

[0035] Variables with a clear progressive, hierarchical, or sequential relationship between categories are defined as ordinal variables and transformed using an ordinal encoder. For example, the severity of coronary artery disease is divided into four levels according to the Z-score: normal, mild, moderate, and severe, coded as 0, 1, 2, and 3, respectively. If clinical symptoms or examination results are described as negative, mild, moderate, and severe, they are also coded in ascending order of their medical severity.

[0036] By following the steps above, different types of categorical variables can be converted into numerical features suitable for model training while preserving the true meaning of clinical variables, thus avoiding misjudgment of the magnitude relationship between variables due to incorrect encoding.

[0037] ④ Class Imbalance Handling: To address the issue of a small number of moderate and severe coronary artery disease cases in Kawasaki disease clinical data, the data acquisition and preprocessing module 100 only uses the SMOTE (Synthetic Minority Over-sampling Technique) or SMOTE for Nominal Continuous methods from the imbalanced-learn library to perform minority class oversampling within the training set.

[0038] Specifically, firstly, the number of samples in the training set for the four categories of normal, mild, moderate, and severe coronary artery disease was counted to identify moderate and severe coronary artery disease cases as minority samples. Then, after missing value imputation, standardization, and category encoding were performed on the training set, new synthetic minority samples were generated in the feature space through random interpolation based on the nearest neighbor relationships of the minority samples. For continuous variables, numerical interpolation was used; for categorical variables such as gender and IVIG resistance, the SMOTENC method was used to specify the categorical variable index to maintain discrete values.

[0039] Oversampling is performed only on the training set or cross-validation training set. Oversampling is not performed on the internal validation set, external validation set, or actual clinical prediction cases to avoid data leakage that could lead to an overestimation of the model's generalization ability. After the synthetic samples are generated, a medical rationality check is performed to remove samples whose laboratory indicators exceed clinically reasonable ranges or whose categorical variables have abnormal values. Finally, the synthetic samples are merged with the original training samples for subsequent model training.

[0040] In this embodiment, the feature filtering module 200 is used to perform joint feature filtering on the training sample set using multiple feature filtering algorithms based on different principles, and the intersection of the filtering results of each algorithm is taken to obtain the core prediction feature set.

[0041] like Figure 2 As shown, the feature filtering module 200 specifically includes: a candidate feature matrix construction unit 210, a first filtering unit 220, a second filtering unit 230, a third filtering unit 240, and an intersection determination unit 250.

[0042] Furthermore, the candidate feature matrix construction unit 210 is used to construct a candidate feature matrix containing demographic features, clinical manifestations, laboratory indicators, and derived inflammatory indicators, using the preprocessed training set as input.

[0043] Specifically, the training set, after missing value imputation, standardization, categorical encoding, and class imbalance processing, is used as the input data for feature selection. Let the feature matrix of the training set be X, and the outcome label be Y, where X includes candidate variables such as demographic characteristics, clinical manifestations, laboratory indicators, and derived inflammatory markers, and Y represents the severity grading of coronary artery lesions, including four categories: normal, mild, moderate, and severe. For a four-class classification task, direct selection using multi-class labels can be employed, or a one-to-many strategy can be used, with normal vs. abnormal, mild vs. non-mild, moderate vs. non-moderate, and severe vs. non-severe as separate selection tasks.

[0044] Furthermore, the first screening unit 220 is used to perform feature selection on the candidate feature matrix using a random forest-based feature selection algorithm, and output a first candidate feature set. Preferably, the random forest-based feature selection algorithm is a random forest combined with recursive feature elimination (RF-RFE) algorithm.

[0045] The training set feature matrix X is input into the RF-RFE algorithm, using a random forest classifier as the basic evaluation model. First, a random forest model is trained using all candidate features, and the candidate features are ranked according to their importance score in the random forest (based on the reduction in Gini impurity or the decrease in out-of-bag accuracy). Then, one or more features with the lowest importance scores are removed. Subsequently, the random forest model is retrained on the remaining features, and the feature importance scores are calculated again. This process is repeated until a preset range of feature counts is reached or model performance no longer improves. In each round of recursive elimination, 10-fold cross-validation is used to evaluate model performance, and metrics such as MCC (Matthews Correlation Coefficient) and ROC-AUC are used as selection criteria to finally determine the optimal feature subset under the RF-RFE algorithm. In this embodiment, the RF-RFE algorithm selects 16 candidate features.

[0046] Furthermore, the second screening unit 230 is used to perform feature selection on the candidate feature matrix using a regularized regression-based feature selection algorithm, and output a second candidate feature set. Preferably, the regularized regression-based feature selection algorithm is a LASSO regression combined with recursive feature elimination (Lasso-RFE) algorithm.

[0047] The training set feature matrix X is input into the Lasso-RFE algorithm, using a Logistic Regression model with L1 regularization or a LASSO model as the base evaluator. First, L1 regularization compresses the feature coefficients, bringing the coefficients of features with low predictive contributions close to zero. Then, the importance of features is ranked according to the absolute value of their regression coefficients. Subsequently, a recursive feature elimination method is used, deleting one or more features with small absolute coefficient values ​​and low predictive contributions in each round, and retraining the model on the remaining features. During the recursive deletion process, 10-fold cross-validation is used to evaluate model performance under different numbers of features, using MCC, ROC-AUC, or the average cross-validation score as evaluation metrics. The subset of features with better performance and fewer features is selected as the Lasso-RFE selection result. In this embodiment, the Lasso-RFE algorithm selects 18 candidate features.

[0048] Furthermore, the third screening unit 240 is used to perform feature selection on the candidate feature matrix using a feature selection algorithm based on statistical tests, and output a third candidate feature set. Preferably, the feature selection algorithm based on statistical tests is the Boruta algorithm.

[0049] The training set feature matrix X is input into the Boruta algorithm, using a random forest classifier as the basic evaluation model. First, the original features are randomly shuffled to generate the same number of shadow features as the original features. Then, the original features and shadow features are jointly input into the random forest model for training, calculating the importance score for each original and shadow feature. Subsequently, the importance score of each original feature is compared with the highest importance score among the shadow features. If the importance score of an original feature is significantly higher than the highest importance score of a shadow feature, the feature is determined to be important; if it is significantly lower, it is determined to be unimportant; if the difference is unclear, the process proceeds to the next iteration. After a preset maximum number of iterations or after all features have been evaluated, the feature set selected by the Boruta algorithm is output. In this embodiment, the Boruta algorithm selects 19 candidate features.

[0050] Furthermore, the intersection determination unit 250 is used to take the intersection of the first candidate feature set, the second candidate feature set and the third candidate feature set, and determine the features selected by multiple algorithms at the same time as the core prediction features.

[0051] Specifically, let the output of the first screening unit 220 be S1 (RF-RFE screening result), the output of the second screening unit 230 be S2 (Lasso-RFE screening result), and the output of the third screening unit 240 be S3 (Boruta screening result). The intersection determination unit 250 takes the intersection of the three feature sets, i.e., S = S1∩S2∩S3, and determines the variables that are simultaneously screened by the three algorithms as the core prediction features.

[0052] In this embodiment, Lasso-RFE selects 18 features, RF-RFE selects 16 features, and Boruta selects 19 features. The intersection of the results from these three algorithms is used to determine the final 8 core prediction features. These 8 core prediction features are then substituted into the subsequent model training process, and the model performance is evaluated on both internal and external validation sets. If the model built based on these 8 core features maintains good MCC, F1-score, ROC-AUC, and PR-AUC, it indicates that the selected features have good predictive value and generalization ability, and can be used as the final input features of the system of this invention.

[0053] Following the above joint screening, the final eight core predictive features identified include: IVIG resistance, gender (male), platelet count (PLT), serum albumin (ALB), platelet-lymphocyte ratio (PLR), total bilirubin (TBIL), CALLY index, and C-reactive protein (CRP). These eight core predictive features are then input into the subsequent Stacking ensemble learning prediction module 300 to construct a four-level stratified prediction model for the severity of coronary artery lesions in Kawasaki disease.

[0054] In this embodiment, the Stacking ensemble learning prediction module 300, using the core prediction feature set as input, includes a first base learner layer 330 and a second meta-learner layer 340. The first base learner layer 330 performs multiple binary classification predictions on the severity of coronary artery lesions and outputs the probability of each binary classification prediction. The second meta-learner layer 340 fuses the probability of each binary classification prediction into a four-level severity classification prediction result and outputs it.

[0055] like Figure 3 As shown, the Stacking ensemble learning prediction module 300 includes a severity label construction unit 310, a task transformation unit 320, a first-layer base learner layer 330 (including a hyperparameter optimization unit and an optimal base learner selection unit), a second-layer meta-learner layer 340 (including a probability vector generation unit, a training data generation unit, a meta-learner training unit, and a probability normalization output unit), and an ordered severity probability correction unit 350.

[0056] Furthermore, the severity label construction unit 310 is used to calculate the coronary artery Z-value based on the coronary ultrasound examination results of the child, and generate a corresponding four-level severity label.

[0057] First, the Z-score of the coronary arteries is calculated based on the results of the child's coronary ultrasound examination. Specifically, the ultrasound physician measures the internal diameter of the left main coronary artery (LMCA), left anterior descending artery (LAD), left circumflex artery (LCX), and the proximal and mid segments of the right coronary artery (RCA-prox and RCA-mid), while simultaneously recording the child's height and weight. The system calculates the body surface area based on height and weight, and then, using the appropriate age or body surface area correction formula, calculates the Z-score for each coronary artery segment.

[0058] The formula for calculating body surface area is: ; BSA represents body surface area, with units of 1000 ppm. H represents the child's height in cm; W represents the child's weight in kg. This represents the square root operation.

[0059] The formula for calculating the coronary Z-score is: ; Where Z represents the Z-value of a certain coronary artery segment; This indicates the actual inner diameter of the coronary artery segment as measured by ultrasound; This represents the normal predicted mean inner diameter of the corresponding coronary artery segment under the same body surface area conditions. This represents the prediction standard deviation of the coronary artery segment under the same body surface area conditions.

[0060] After obtaining the child's body surface area, the system substitutes the body surface area into the preset coronary artery diameter reference equation to obtain the predicted average internal diameters of the left main coronary artery, left anterior descending artery, left circumflex artery, and the proximal and mid-segments of the right coronary artery. and the predicted standard deviation Subsequently, the actual inner diameter of each segment was measured. Substituting these values ​​into the Z-value formula above, we obtain the Z-values ​​for each coronary artery segment.

[0061] Finally, the Z-scores of all coronary artery segments are compared, and the maximum value is used as the final criterion for determining the severity of the child's coronary artery disease. If the maximum Z-score is less than 2.5, it is considered normal; if 2.5 ≤ maximum Z-score < 5, it is considered mild coronary artery disease; if 5 ≤ maximum Z-score < 10, it is considered moderate coronary artery disease; if the maximum Z-score is ≥ 10, or the coronary artery diameter meets the criteria for severe aneurysm, it is considered severe coronary artery disease.

[0062] Subsequently, the four severity levels were assigned numerical labels: 0 for normal, 1 for mild, 2 for moderate, and 3 for severe, forming a multi-category label Y. Each child was assigned to only one highest severity category. If different coronary artery segments corresponded to different levels based on Z-scores, the highest level was used as the final label.

[0063] Furthermore, the task conversion unit 320 is used to convert the four severity labels into four one-to-many binary classification tasks.

[0064] Considering the significant class imbalance in the severity of coronary artery lesions in Kawasaki disease, especially the relatively small number of moderate and severe cases, directly performing a single four-class classification model would easily bias the model towards the normal or mild category, leading to missed classifications of moderate and severe cases. Therefore, the task transformation unit 320 transforms the four-level classification task into four one-to-many binary classification tasks.

[0065] The four one-to-many binary classification tasks specifically include: Task A: Normal vs. Abnormal. "Normal" refers to children with a coronary Z-score < 2.5; "abnormal" refers to all children with coronary artery disease other than normal, including children with mild, moderate, and severe coronary artery disease, i.e., children with a Z-score ≥ 2.5. Normal cases were labeled as positive (1), and mild, moderate, and severe cases were labeled as negative (0).

[0066] Task B: Mild vs. Non-mild. "Mild" refers to children with a coronary Z-score of 2.5 ≤ Z-score < 5; "Non-mild" refers to all children other than those with mild cases, including normal, moderate, and severe cases. Mild cases were labeled as positive (1), and normal, moderate, and severe cases were labeled as negative (0).

[0067] Task C: Moderate vs. Non-moderate. "Moderate" refers to children with a coronary Z-score of 5 ≤ Z-score < 10; "Non-moderate" refers to all children other than those with moderate cases, including normal, mild, and severe cases. Moderate cases were labeled as positive (1), and normal, mild, and severe cases were labeled as negative (0).

[0068] Task D: Severe vs. Non-severe. "Severe" refers to children with a coronary artery Z-score ≥10 or a coronary artery diameter meeting the criteria for severe aneurysm; "non-severe" refers to all children other than those with severe aneurysm, including normal, mild, and moderate cases. Severe cases were labeled as positive (1), and normal, mild, and moderate cases were labeled as negative (0).

[0069] Furthermore, the first base learner layer 330 is trained and predicted for the four one-to-many binary classification tasks mentioned above.

[0070] The first base learner layer 330 includes multiple candidate base learners, a hyperparameter optimization unit, and an optimal base learner selection unit.

[0071] ① Candidate base learners The first base learner layer 330 includes multiple candidate base learners, which are selected from any combination of AdaBoost, CatBoost, Decision Tree, ExtraTrees, GBDT, KNN, LightGBM, RandomForest, SVM, and XGBoost. Each candidate base learner uses features from the core prediction feature set as input variables and the label corresponding to the binary classification task as output variables.

[0072] The principles of each candidate basis learner are as follows: AdaBoost (Adaptive Boosting) is an adaptive boosting algorithm. Its basic idea is to train multiple weak classifiers sequentially, increasing the weight of misclassified samples from the previous round in each training iteration, thus making subsequent classifiers focus more on samples that are difficult to identify. Finally, the multiple weak classifiers are weighted and combined according to their classification performance to form a comprehensive classifier.

[0073] CatBoost is a machine learning model based on gradient boosting decision trees. It can handle categorical variables well and reduces the risk of overfitting through an ordered boosting strategy.

[0074] A decision tree is a classification model based on a tree structure. By setting splitting conditions on the input features, it progressively divides the samples into different leaf nodes and outputs the classification result based on the distribution of sample categories in the leaf nodes.

[0075] ExtraTrees (Extremely Randomized Trees) are an ensemble model composed of multiple random decision trees. Compared to ordinary random forests, ExtraTrees introduce stronger randomness in feature selection and split threshold selection, thereby reducing model variance and improving generalization ability.

[0076] GBDT (Gradient Boosting Decision Tree) is an ensemble learning model that reduces the prediction error of the previous model by progressively training multiple decision trees. Each new tree primarily learns the residual information that the previous model failed to accurately predict, thereby continuously improving the overall prediction performance.

[0077] KNN (K-Nearest Neighbors) is a classification model based on sample distance. For a child to be predicted, KNN finds the K samples in the training set that are closest to its features, and determines the probability that the child belongs to the target class based on the class distribution of these neighboring samples.

[0078] LightGBM is a lightweight gradient boosting tree model characterized by fast training speed and high efficiency in processing features. It improves the training efficiency and predictive ability of the model on structured clinical data through histogram-based feature splitting and a leaf-first growth strategy.

[0079] Random forest is an ensemble classification model consisting of multiple decision trees. It constructs multiple different decision trees by randomly sampling training samples and features, and combines the prediction results of each decision tree to obtain the final classification probability.

[0080] SVM (Support Vector Machine) is a model that achieves classification by finding the optimal classification hyperplane. Its goal is to construct the maximum margin classification boundary between samples of different classes and can handle nonlinear classification problems through kernel functions.

[0081] XGBoost (Extreme Gradient Boosting) is an efficient ensemble learning model based on gradient boosting trees. It reduces the risk of overfitting and improves the model's predictive performance through regularization terms, column sampling, row sampling, and iterative boosting mechanisms.

[0082] The aforementioned 10 models have different learning mechanisms: decision trees, random forests, and ExtraTrees focus on tree-like rule splitting; AdaBoost, GBDT, LightGBM, XGBoost, and CatBoost focus on boosting learning; KNN focuses on sample similarity judgment; and SVM focuses on classification boundary construction. By introducing different types of base learners, this invention can learn the relationship between the clinical characteristics of children with Kawasaki disease and the severity of coronary artery lesions from multiple perspectives.

[0083] ② Hyperparameter optimization unit The hyperparameter optimization unit employs a Bayesian optimization method, using the combined scores of MCC, F1-score, and PR-AUC as the objective function to search for and optimize the hyperparameters of each candidate base learner.

[0084] To avoid model performance instability caused by manually setting parameters based on experience, this invention employs a combination of Bayesian optimization and manual boundary correction for hyperparameter optimization. The specific steps are as follows: The first step is to pre-define the hyperparameters to be optimized and their search range for each type of base learner. For example, the parameters to be optimized for Random Forest and ExtraTrees include the number of trees (n_estimators), maximum tree depth (max_depth), minimum number of leaf node samples (min_samples_leaf), and number of candidate features per split (max_features); the parameters to be optimized for CatBoost, LightGBM, XGBoost, and GBDT include the learning rate (learning_rate), tree depth (max_depth), number of iterations (n_estimators), number of leaf nodes (num_leaves), regularization parameters (reg_alpha, reg_lambda), and subsample ratio; the parameters to be optimized for SVM include the penalty coefficient C, kernel type (kernel), and kernel parameter (gamma); the parameters to be optimized for KNN include the number of nearest neighbors (n_neighbors), distance metric, and weights; and the parameters to be optimized for AdaBoost include the number of weak learners (n_estimators), learning rate (learning_rate), and the depth of the base classifier (base_estimator's max_depth).

[0085] The second step involves randomly selecting several sets of initial parameter combinations within the defined search space, training corresponding base learners, and calculating model performance through 10-fold cross-validation.

[0086] The third step involves establishing a surrogate model between hyperparameters and model performance, using the combined score of MCC, F1-score, and PR-AUC as the objective function. The formula for calculating the combined score is: Overall score = 0.4 × MCC + 0.3 × F1 score + 0.3 × PR-AUC Among them, MCC is used to evaluate the overall classification ability of the model under class imbalance, F1-score is used to balance positive prediction accuracy and recall, and PR-AUC (Precision-Recall Area Under Curve) is used to evaluate the ability to identify minority cases. Since there are fewer moderate and severe CAL cases, PR-AUC and MCC are more suitable as optimization criteria than accuracy alone.

[0087] The fourth step is to select the next set of hyperparameters that are most likely to improve the objective function based on the Bayesian optimized acquisition function (such as EI, Expected Improvement), and continue to train and evaluate the model.

[0088] Fifth, repeat the above search process until the preset maximum number of iterations is reached, or the overall score no longer improves significantly in multiple consecutive iterations.

[0089] The sixth step involves manually correcting the optimal parameter combination obtained from Bayesian optimization, eliminating parameters that clearly lead to overfitting or do not meet clinical generalization requirements. For example, when the tree model depth is too large and the training set performance is significantly higher than the cross-validation performance, the tree depth is reduced or the minimum number of leaf node samples is increased; when the Boosting model has undergone too many iterations and the validation performance no longer improves, an early stopping strategy is implemented or the number of iterations is reduced.

[0090] Step 7: Fix the random seed, save the optimal hyperparameter combination and cross-validation performance results of each base learner for each task, and ensure that the model training results are reproducible.

[0091] Furthermore, this invention does not prioritize accuracy as the primary optimization objective. Instead, it introduces a minority-sensitive optimization objective for identifying the minority class of moderate to severe coronary artery disease (CAL). Specifically, in the tasks of moderate vs. non-moderate and severe vs. non-severe, the weights of MCC, recall, and PR-AUC are increased. When the AUCs of the two models are similar, the model with higher recall and lower false negative rate for moderate to severe CAL is prioritized. This optimization approach does not simply pursue overall accuracy but is designed around the technical problem of "moderate to severe cases being rare but posing a greater clinical threat" in the prediction of the severity of Kawasaki disease CAL, thereby reducing the risk of missing severe coronary artery lesions.

[0092] ③ Optimal basis learner selection unit The optimal base learner selection unit performs optimal base learner selection for each binary classification task.

[0093] In each binary classification task, 10 candidate base learners were trained and cross-validated, and evaluation metrics for each model were calculated, including ROC-AUC, PR-AUC, MCC, F1-score, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). For each model, the mean performance and standard deviation of its 10-fold cross-validation were recorded to evaluate model stability.

[0094] The filtering rules include the following five priorities: First priority rule: Among multiple candidate base learners, prioritize the model with the highest MCC. For the same binary classification task, if the average MCC of a certain model is higher than that of other models, and the difference is ≥0.02, then directly prioritize that model as the candidate optimal base learner.

[0095] The second priority rule is as follows: When the MCC difference among multiple candidate base learners is less than 0.02 (i.e., their MCCs are close), further compare PR-AUC and F1-score. In practice, first compare PR-AUC. If a model's PR-AUC is ≥0.02 higher than other models, that model is selected first. If the PR-AUC difference is also <0.02, then continue comparing F1-score. If a model's F1-score is ≥0.02 higher, that model is selected first. PR-AUC better reflects the recognition ability of minority class samples, while F1-score comprehensively reflects precision and recall.

[0096] The third priority rule: For moderate vs. non-moderate tasks and severe vs. non-severe tasks, when the difference in overall performance metrics is less than 0.02, the model with the highest recall rate is selected first. Here, "the difference in overall performance metrics is less than 0.02" means that the difference between the two models in the three main metrics—MCC, PR-AUC, and F1-score—is all < 0.02. The specific implementation steps are as follows: First, select models whose differences in MCC, PR-AUC, and F1-score are all < 0.02; second, compare the recall rates of these models for moderate or severe positive cases; third, if the recall rate of one model is ≥ 0.02 higher than the other models, then select that model first; finally, if the difference in recall rate is still < 0.02, proceed to the next step of stability comparison.

[0097] The fourth priority rule: When the difference in all performance metrics is less than 0.02, the model with the smallest cross-validation standard deviation is selected first. In practice, the standard deviations of MCC, PR-AUC, and F1-score in the 10-fold cross-validation of each model are calculated. If a model has a smaller standard deviation for its main metrics, and the difference in standard deviations is ≥0.01, then the model is considered to have less fluctuation and better stability across different training folds, and this model is selected first. If the difference in standard deviations of multiple models is <0.01, then their stability is considered similar, and the next step is to compare model complexity.

[0098] Fifth priority rule: When performance and stability are similar, prioritize models with fewer parameters and shorter prediction time. "Similar performance and stability" means: 1) the differences in MCC, PR-AUC, F1-score, and recall between models are all <0.02; 2) the differences in the standard deviations of the main 10-fold cross-validation metrics are all <0.01. In this case, further compare model complexity and deployment ease, including the number of model parameters, prediction time, model file size, and ease of embedding in web applications. If a model has shorter prediction time, smaller model file size, and simpler parameter structure, then that model is prioritized as the final base learner.

[0099] By using the above selection rules, this invention can ensure the model's predictive performance while also taking into account minority class identification capabilities, cross-validation stability, and ease of clinical deployment.

[0100] In this embodiment, according to the above selection rules, the optimal base learners for the four binary classification tasks are determined as follows: ExtraTrees model is selected for normal vs. abnormal tasks; CatBoost model is selected for mild vs. moderate tasks; AdaBoost model is selected for moderate vs. moderate tasks; and AdaBoost model is selected for severe vs. mild tasks. These four optimal models are used as the base learners for the first layer of the Stacking ensemble model.

[0101] This invention does not use the same model across all severity levels. Instead, it selects the optimal base learner for each of the four tasks: normal, mild, moderate, and severe. Since the clinical indicators for different severity levels of CAL (Critical Ailment Scale) vary, a single model may not be suitable for all levels simultaneously. This invention employs a "task-specific optimal base learner" strategy, ensuring that each severity level is identified by the model best suited to its feature distribution, and then fused by a second-layer meta-learner, thereby improving overall prediction stability and minority class identification ability.

[0102] Furthermore, the second meta-learner layer 340 includes a probability vector generation unit, a training data generation unit, a meta-learner training unit, and a probability normalization output unit.

[0103] ①Probability Vector Generation Unit The probability vector generation unit is used to input the core prediction features of each child into the four optimal base learners to obtain the prediction probabilities of the normal category, mild category, moderate category, and severe category, and to combine the four prediction probabilities into a four-dimensional probability vector as secondary features.

[0104] Specifically, the eight core features of each child are input into the four optimal base learners mentioned above to obtain the positive prediction probabilities for four binary classification tasks, which are denoted as follows: The predicted probability that the child belongs to the normal category; The predicted probability that the child belongs to the mild category; The predicted probability that the child belongs to the moderate category; The predicted probability that the child belongs to the severe category.

[0105] The four predicted probabilities mentioned above together constitute a new secondary feature vector. .

[0106] ② Training data generation unit The training data generation unit is used to generate meta-learner training data from the training set using K-fold cross-validation.

[0107] To avoid overfitting caused by the first-layer model directly predicting on the training set, this invention uses cross-validation to generate training data for the meta-learner. Specifically, K-fold cross-validation (K=10 in this embodiment) is performed on the training set. In each fold, K-1 folds are used to train the first-layer base learner, and the remaining 1 fold is used to generate the predicted probability of the sample in that fold. After repeating K times, all training samples obtain four predicted probabilities output by the base learner that has not seen the sample before. This four-dimensional probability vector is used as the input feature of the meta-learner, and the original four-level CAL severity labels are used as the output labels of the meta-learner.

[0108] In this way, the meta-learner learns the predicted distribution of unknown samples by the first-layer base learner, rather than the memorized results of the training samples, thereby reducing the risk of overfitting in the Stacking model.

[0109] ③ Meta-learner training unit The meta-learner training unit is used to train a four-class fusion model by using a logistic regression model as the second-layer meta-learner, the four-dimensional probability vector as the input feature, and the true severity label of the child's coronary artery lesions as the output label.

[0110] Specifically, Logistic regression is used as the second-layer meta-learner. The meta-learner uses the probability vectors output by the four base learners. As input, the true severity label Y of the child's chronic arterial disease (CAL) is used as output to train a four-class fusion model. The Logistic Regression Meta-Learner outputs the final fusion probabilities of the four categories, corresponding to normal, mild, moderate, and severe, respectively.

[0111] The reason for choosing Logistic Regression as the meta-learner is that its parameter structure is relatively simple, which can reduce the risk of overfitting in the second layer fusion; at the same time, it can weight and fuse the probabilities output by different models in the first layer, so that the final result has good interpretability and deployability.

[0112] ④ Probability Normalized Output Unit The probability normalization output unit is used to output the fusion probability of each child belonging to one of the four categories: normal, mild, moderate, and severe, and the category with the highest probability is used as the final predicted severity.

[0113] The meta-learner outputs the final probability that each child belongs to one of four categories: normal, mild, moderate, or severe, denoted as [missing information]. , , , The system normalizes the probabilities of the four categories so that their sum is 1. The final classification result is determined according to the principle of maximizing probability, i.e., selecting... , , , The category corresponding to the highest median score is used as the final predicted severity. For example, if... =0.18、 =0.30、 =0.42、 If the value is 0.10, the system will output "Highest risk of moderate coronary artery disease".

[0114] Furthermore, the Stacking ensemble learning prediction module 300 also includes an ordered severity probability correction unit 350, which performs the following correction steps before the probability normalization output unit outputs the final prediction result: Because the severity of CAL progresses from normal, mild, moderate to severe, the system performs adjacent-level consistency correction on the probabilities of the four categories before outputting the final category.

[0115] The specific steps are as follows: The first step is to calculate the probability difference between the category with the highest final probability and its adjacent categories. For example, if the highest category is mild, then calculate the probability difference between mild and normal, and between mild and moderate.

[0116] The second step involves marking the case as a "borderline risk case" when the difference between the highest probability category and the adjacent higher severity category is less than a preset threshold. For example, if the probability of mild severity is 0.41 and the probability of moderate severity is 0.38, and the difference between the two is less than 0.05, the system outputs a mild prediction result while prompting "Mild-moderate borderline risk, ultrasound follow-up verification is recommended."

[0117] Third, when the moderate or severe predicted probability exceeds the safety threshold, the system will still output a moderate to severe risk warning, even if it is not the highest probability category. For example, if the severe probability does not reach the highest but exceeds the preset threshold, the system will prompt "There is a severe CAL risk signal, and it is recommended to pay close attention."

[0118] The fourth step is for the system to output a four-level probability distribution simultaneously, instead of just a single category, so that clinicians can assess the uncertainty of the model's predictions.

[0119] This probability correction mechanism incorporates the order of medical severity into the Stacking fusion process, avoiding treating the four levels of CAL as completely independent and disordered categories, and reducing the probability of moderate to severe cases being misclassified as low-risk.

[0120] Furthermore, the Stacking ensemble learning prediction module 300 constructs the Stacking ensemble model according to the following steps: The first step is to input the training set's eight core prediction features and CAL four-level severity labels.

[0121] The second step is to convert the four-level labels into four one-to-many binary classification task labels.

[0122] The third step is to train 10 candidate base learners for each binary classification task.

[0123] The fourth step involves using Bayesian optimization combined with 10-fold cross-validation to determine the optimal hyperparameters for each base learner.

[0124] The fifth step is to select the optimal base learner for each task based on MCC, F1-score, PR-AUC, recall, and model stability.

[0125] The sixth step is to use cross-validation to generate the training set prediction probabilities of the four optimal base learners, forming the meta-learner training matrix.

[0126] The seventh step is to input the predicted probability matrix into the Logistic Regression Meta-learner to train the four-class fusion model.

[0127] The eighth step is to save the four first-layer base learners and the second-layer Logistic regression meta-learner together after the model training is completed, forming a complete Stacking ensemble prediction model.

[0128] The ninth step involves inputting the eight core features of the child to be predicted into the model, which then sequentially passes through the first layer of base learner probability output, the second layer of meta learner fusion, and ordered severity probability correction, ultimately outputting four levels of prediction results: normal, mild, moderate, and severe, along with their corresponding probabilities.

[0129] After the model is trained, its performance is evaluated using both internal and external validation sets.

[0130] Internal Validation Set Evaluation: The trained Stacking ensemble model was applied to the internal validation set. The internal validation set was not involved in model training, oversampling, feature selection, or hyperparameter optimization; it was used solely for model performance evaluation. For each child in the internal validation set, eight core predictive features were input, and the model output a four-level CAL severity prediction result. Subsequently, the prediction results were compared with the actual ultrasound Z-score grading results to construct a confusion matrix, and sensitivity, specificity, positive predictive value, negative predictive value, F1-score, MCC, ROC-AUC, and PR-AUC were calculated. For the four levels—normal, mild, moderate, and severe—the above indicators were calculated using a one-to-many approach, and macro-average and weighted-average indicators were calculated to evaluate the overall model performance.

[0131] External Validation Set Evaluation: To evaluate the model's generalization ability, the trained and parameter-fixed Stacking ensemble model was directly applied to external validation cohorts from three other hospitals. The external validation data underwent only the same missing value handling, standardization, and class encoding as the training set; the model was not retrained, features were not reselected, and SMOTE oversampling was not re-performed. For each child in the external validation set, eight core features were input, and the predicted severity of CAL (Cognitive Ailment) at four levels and its probability distribution were output. Subsequently, the predicted results were compared with the true Z-scores, and the confusion matrix, sensitivity, specificity, positive predictive value, negative predictive value, F1-score, MCC, ROC-AUC, and PR-AUC of the external validation set were calculated. In this example, the model maintained good MCC, F1-score, ROC-AUC, and PR-AUC in both the internal and external validation sets, indicating that the Stacking ensemble model has good stability and cross-center generalization ability.

[0132] Comparative Validation with Single Models: To further demonstrate the effectiveness of the Stacking ensemble model, this invention compares the Stacking model with various single base learners. Specifically, AdaBoost, CatBoost, Decision Tree, ExtraTrees, GBDT, KNN, LightGBM, Random Forest, SVM, and XGBoost are treated as independent models, trained and evaluated on the same training set, the same core features, and the same validation set. The MCC, F1-score, ROC-AUC, and PR-AUC of each model are compared on the internal and external validation sets. In this example, the Stacking model outperforms most single models on key evaluation metrics, especially exhibiting higher MCC, F1-score, or PR-AUC in the identification of moderate and severe CAL lesions, demonstrating that the two-layer ensemble structure can integrate the advantages of different models and improve the identification ability of a minority of severe coronary artery lesions.

[0133] Model output: Finally, the Stacking ensemble model outputs the following: First, the predicted probability of the child belonging to four levels: normal, mild, moderate, and severe CAL; Second, the predicted category of CAL severity with the highest probability; Third, a moderate or severe risk indication; Fourth, a case indication of adjacent level boundaries; Fifth, the model input and prediction results used for the subsequent SHAP interpretation module.

[0134] In this embodiment, the interpretability analysis module 400 is used to receive the prediction results output by the Stacking ensemble learning prediction module, calculate the contribution value of each feature in the core prediction feature set to the prediction results, and generate visual explanation information.

[0135] Preferably, the interpretability analysis module 400 is implemented based on the SHAP (SHapley Additive exPlanations) method. For example... Figure 4 As shown, the interpretability analysis module 400 includes: a SHAP value calculation unit 410, a global feature importance analysis unit 420, an individualized interpretation unit 430, and a visualization output unit 440.

[0136] Furthermore, the SHAP value calculation unit 410 is used to input the trained Stacking ensemble model, training set background data, and core prediction features into the SHAP interpreter to calculate the SHAP value of each feature in each sample for the prediction result.

[0137] The SHAP value represents the contribution of a particular feature to the model's prediction result. Its basic expression is: ; in, The model represents a sample of patients. The predicted output value; This indicates the patient's eight core characteristics; This represents the baseline prediction of the model, i.e., the average predicted output of the training set samples; Indicates the first The contribution value of each feature to the prediction result; Indicates the feature number, ; This represents the sum of the contribution values ​​of the eight characteristics. If... This indicates that the feature drives the model to predict the corresponding severity; if This indicates that the feature reduces the likelihood of the model predicting the corresponding severity. The larger the value, the greater the influence of the feature on the prediction result.

[0138] In practice, the SHAP value calculation unit 410 uses either the KernelSHAP or TreeSHAP algorithm to calculate the SHAP value. For tree-based models (such as ExtraTrees, CatBoost, and AdaBoost), the TreeSHAP algorithm can accurately calculate the SHAP value in polynomial time, resulting in higher computational efficiency. The trained Stacking ensemble model, training set background data (a subset of 100-500 samples randomly selected from the training set), and eight core prediction features are input into the SHAP interpreter to calculate the SHAP value of each feature for each sample in relation to the prediction result.

[0139] Furthermore, the global feature importance analysis unit 420 is used to average the absolute values ​​of each feature of all samples in the training set or validation set to obtain the overall importance of each feature, and output the feature importance ranking in descending order of overall importance.

[0140] Specifically, the global feature importance analysis unit 420 calculates the overall importance of each feature according to the following formula: ; in, Indicates the first The overall importance of each feature; This represents the total number of samples involved in the calculation; Indicates the first One sample; Indicates the first In the nth sample The SHAP value of each feature; This represents the absolute value of the SHAP value. According to... Sort the features from largest to smallest to obtain the importance ranking of the 8 core features.

[0141] By analyzing the global importance of features, clinicians can understand which features contribute most to predicting the severity of CAL in the general population, thereby deepening their understanding of disease risk factors.

[0142] Furthermore, the individualized interpretation unit 430 calculates the SHAP value of each core feature for a single child to be predicted and sorts them according to the size of the contribution value, outputting the main factors that drive the prediction of the child to the corresponding severity and the direction of their influence.

[0143] For a single child to be predicted, the system calculates the SHAP values ​​of its eight core features and sorts them according to their contribution values, showing the main factors that drive the prediction of the child as having normal, mild, moderate, or severe coronary artery disease. For example, when the SHAP values ​​corresponding to IVIG resistance, elevated CRP, and decreased ALB are positive, it indicates that these factors drive the model to output a higher risk of CAL; when the SHAP value of a certain indicator is negative, it indicates that the indicator reduces the corresponding risk.

[0144] By interpreting the results individually, clinicians can understand the reasons behind the predicted outcomes for each specific child, increase their confidence in the predictions, and provide a basis for individualized treatment decisions.

[0145] Furthermore, the visualization output unit 440 is used to generate feature importance bar charts, SHAP waterfall plots, and SHAP force plots. The feature importance bar charts are used to display the importance of each feature in the overall population; the waterfall plots or force plots are used to show how each feature of a single patient affects the prediction result from the baseline value to the final predicted value.

[0146] The waterfall plot uses arrows and bars to show the complete path of change from the baseline prediction to the final prediction. Each feature corresponds to a bar, and the color indicates whether the feature contributes positively or negatively to the prediction result. It aims to illustrate the push-pull effect of each feature on the prediction result in the form of a force field, making the causes of the prediction result immediately clear.

[0147] In this embodiment, the Web application deployment module 500 is used to encapsulate the core prediction feature set, the Stacking integrated learning prediction module, and the interpretability analysis module into a lightweight Web application, providing a feature input interface, a prediction result display interface, and an interpretable visualization display interface.

[0148] Preferably, the Web application deployment module 500 is built based on the Streamlit framework. For example... Figure 5 As shown, the Web application deployment module 500 includes: an input interface unit 510, a background prediction unit 520, a result display unit 530, and a result export unit 540.

[0149] Furthermore, the input interface unit 510 is used to provide input interfaces for each feature in the core prediction feature set. The input interface supports binary selection input for IVIG resistance (yes / no), binary selection input for gender (male / female), and numerical input for platelet count (PLT), serum albumin (ALB), platelet-lymphocyte ratio (PLR), total bilirubin (TBIL), CALLY index, and C-reactive protein (CRP). Each numerical input box is configured with a reasonable numerical range validation; for example, the platelet count range is set to... Serum albumin range is set to 10-60 g / L, C-reactive protein range is set to 0-500 mg / L, etc., to avoid abnormal prediction results caused by input errors.

[0150] Furthermore, the background prediction unit 520 is used to embed the Stacking ensemble learning prediction module and the interpretability analysis module into the web application backend. In response to a prediction trigger operation after input completion (e.g., clicking the "Predict" button), the Stacking ensemble learning prediction module is invoked to perform prediction calculations, and the interpretability analysis module is invoked to generate corresponding visual explanation information. After receiving the input data, the background prediction unit automatically performs standardization processing on the input data according to the same rules as the training set, and then invokes the loaded Stacking ensemble model for inference, outputting the predicted severity of the child's CAL (normal, mild, moderate, severe) within 1-2 seconds.

[0151] Furthermore, the results display unit 530 is used to simultaneously display the predicted probability distribution of the four severity levels, the predicted category of the final severity, the SHAP waterfall plot, and the feature contribution ranking. Specifically, the results display unit shows the predicted probability distribution of the four categories in a bar chart or pie chart, highlights the final predicted category, and displays the corresponding SHAP waterfall plot or force plot for the child below, so that doctors can see both "what is predicted" and "why this prediction is made".

[0152] Furthermore, the results export unit 540 supports the export of prediction results and interpretability analysis reports for inclusion in the patient's medical record. Export formats include PDF and HTML. The report content includes the patient's basic identification information (de-identified), input feature values, four-level prediction probability distribution, final prediction category, SHAP interpretation chart, and generation timestamp, facilitating traceability of the diagnosis and treatment process.

[0153] The web application is deployed on a cloud server, generating a publicly accessible URL. Doctors do not need to install any software; they can access the application simply through a browser to perform real-time bedside predictions in the clinic or ward, meeting the needs of immediate clinical assessment.

[0154] Application testing was conducted in pediatric clinical settings at four partner hospitals to optimize the ease of use of the input interface and the model's response speed, ensuring that the application can adapt to the usage needs of different clinical environments.

[0155] In summary, this invention proposes a predictive system for the severity of coronary artery lesions in Kawasaki disease, the core of which is embodied in the following technical path: First, the data acquisition and preprocessing module 100 performs missing value imputation, standardization, class coding, and class imbalance processing on the multi-center clinical raw data to generate a standardized training sample set, providing a high-quality data foundation for subsequent modeling.

[0156] Secondly, the feature selection module 200 uses three algorithms with different principles—Boruta, LASSO-RFE, and RF-RFE—to perform joint feature selection and take the intersection. From the 38 original parameters, 8 core predictive features (IVIG resistance, gender, PLT, ALB, PLR, TBIL, CALLY index, and CRP) are selected, which greatly reduces the model complexity while ensuring prediction accuracy.

[0157] Furthermore, a two-layer ensemble learning architecture is constructed through the Stacking ensemble learning prediction module 300: The first layer, the base learner layer 330, decomposes the four-level classification task into four one-to-many binary classification tasks, trains and selects the optimal base learners specific to each task (ExtraTrees for normal vs. abnormal tasks, CatBoost for mild vs. non-mild tasks, and AdaBoost for moderate vs. non-moderate and severe vs. non-severe tasks); The second layer, the meta-learner layer 340, generates training data through cross-validation, and then uses Logistic regression to fuse the prediction probabilities of each base learner into a four-level classification result, and introduces an ordered severity probability correction mechanism to achieve four-level hierarchical prediction of the severity of Kawasaki disease coronary artery lesions.

[0158] Then, the interpretability analysis module 400 calculates the contribution value of each feature to the prediction result based on the SHAP method, and generates a global feature importance ranking and an individualized waterfall plot, which makes up for the shortcomings of traditional machine learning models in terms of the interpretability of internal decision logic.

[0159] Finally, the Web application deployment module 500, based on the Streamlit framework, encapsulates all the above modules into a lightweight web application and deploys it in the cloud, enabling clinicians to complete real-time bedside predictions and result interpretations through a browser.

[0160] Through the organic synergy of the above modules, this invention effectively solves the technical problems in the prior art, such as the inability to achieve four-level stratified prediction of the severity of coronary artery lesions, low prediction accuracy of single models, poor interpretability, and high threshold for clinical application, and achieves a balance between prediction accuracy and clinical usability.

[0161] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0162] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0163] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0164] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A system for predicting the severity of coronary artery lesions in Kawasaki disease, characterized in that, include: The data acquisition and preprocessing module is used to acquire raw clinical data of children with Kawasaki disease, and to generate a training sample set after processing the raw clinical data by imputing missing values, standardization, class coding and class imbalance. The feature selection module is used to perform joint feature selection on the training sample set using multiple feature selection algorithms based on different principles, and to obtain the core prediction feature set by taking the intersection of the selection results of each algorithm. The Stacking ensemble learning prediction module is used to take the core prediction feature set as input. The Stacking ensemble learning prediction module includes a first base learner layer and a second meta learner layer. The first base learner layer performs multiple binary classification predictions on the severity of coronary artery lesions and outputs the probability of each binary classification prediction. The second meta learner layer merges the probability of each binary classification prediction into a four-level severity classification prediction result and outputs it. The interpretability analysis module is used to receive the prediction results output by the Stacking ensemble learning prediction module, calculate the contribution value of each feature in the core prediction feature set to the prediction results, and generate visual explanation information. The Web application deployment module is used to encapsulate the core prediction feature set, the Stacking integrated learning prediction module, and the interpretable analysis module into a lightweight Web application, providing feature input interface, prediction result display interface, and interpretable visualization display interface; The Stacking ensemble learning prediction module further includes a task transformation unit, which transforms the four severity levels into four one-to-many binary classification tasks. These four one-to-many binary classification tasks include: Normal vs. abnormal tasks: normal category samples are labeled as positive, and mild, moderate and severe category samples are labeled as negative. For both light and heavy tasks, light category samples are labeled as positive, while normal, moderate, and heavy category samples are labeled as negative. For moderate vs. non-moderate tasks, moderate category samples are labeled as positive, while normal, mild, and severe category samples are labeled as negative. For heavy vs. light tasks, heavy category samples are labeled as positive, while normal, light, and moderate category samples are labeled as negative. The first base learner layer is trained and predicted for the four one-to-many binary classification tasks respectively; The second meta-learner layer includes: The probability vector generation unit is used to input the core prediction features of each child into the training base learners corresponding to the four one-to-many binary classification tasks to obtain the prediction probability of normal category, mild category, moderate category and severe category, and combine the four prediction probabilities into a four-dimensional probability vector as secondary features. The training data generation unit is used to generate meta-learner training data in the training set using K-fold cross-validation. In each fold, the first-layer base learner is trained using K-1 folds, and the remaining 1 fold is used to generate the predicted probability of the sample in that fold. After repeating K times, a four-dimensional probability vector of all training samples is obtained. The meta-learner training unit is used to train a four-class fusion model by using a logistic regression model as the second-layer meta-learner, the four-dimensional probability vector as the input feature, and the true severity label of the child's coronary artery lesions as the output label. The probability normalization output unit is used to output the fusion probability of each child belonging to one of the four categories: normal, mild, moderate, and severe, and the category with the highest probability is used as the final predicted severity.

2. The system for predicting the severity of coronary artery lesions in Kawasaki disease according to claim 1, characterized in that, In the data acquisition and preprocessing module, the class imbalance handling specifically includes: using synthetic minority class oversampling technology to amplify samples only in the training set corresponding to the moderate and severe categories of coronary artery lesion severity, so that the number of samples in the moderate and severe categories tends to be balanced with the number of samples in the majority category. The internal validation set, external validation set, and actual clinical prediction cases are not subjected to the sample amplification.

3. The system for predicting the severity of coronary artery lesions in Kawasaki disease according to claim 1, characterized in that, The feature filtering module specifically includes: The candidate feature matrix construction unit is used to construct a candidate feature matrix containing demographic features, clinical manifestations, laboratory indicators, and derived inflammatory indicators, using the preprocessed training set as input. The first screening unit is used to perform feature screening on the candidate feature matrix using a feature selection algorithm based on random forest, and output a first candidate feature set. The second screening unit is used to perform feature screening on the candidate feature matrix using a feature selection algorithm based on regularized regression, and output a second candidate feature set. The third screening unit is used to perform feature screening on the candidate feature matrix using a feature selection algorithm based on statistical testing, and output a third candidate feature set. The intersection determination unit is used to take the intersection of the first candidate feature set, the second candidate feature set, and the third candidate feature set, and determine the features that are simultaneously selected by multiple algorithms as the core prediction features.

4. The system for predicting the severity of coronary artery lesions in Kawasaki disease according to claim 3, characterized in that, The core predictive features include: IVIG resistance, sex, platelet count, serum albumin, platelet-lymphocyte ratio, total bilirubin, CALLY index, and C-reactive protein.

5. The system for predicting the severity of coronary artery lesions in Kawasaki disease according to claim 1, characterized in that, The Stacking ensemble learning prediction module also includes a severity label construction unit, which is used to calculate the coronary artery Z-score based on the coronary ultrasound examination results of the child. The formula for calculating the coronary artery Z-score is as follows: ； in, This indicates the actual inner diameter of the coronary artery segment as measured by ultrasound. This represents the normal predicted mean inner diameter of this coronary artery segment under the same body surface area conditions. This represents the prediction standard deviation for the coronary artery segment under the same body surface area conditions. The severity label construction unit compares the Z-values ​​of all coronary artery segments and takes the maximum value as the judgment criterion. When the maximum Z-value is less than 2.5, it is judged as normal; when 2.5 ≤ maximum Z-value < 5, it is judged as mild coronary artery disease; when 5 ≤ maximum Z-value < 10, it is judged as moderate coronary artery disease; and when the maximum Z-value ≥ 10, it is judged as severe coronary artery disease, and a four-level severity label is generated accordingly.

6. The system for predicting the severity of coronary artery lesions in Kawasaki disease according to claim 1, characterized in that, The first base learner layer includes multiple candidate base learners, which are selected from any of the following: AdaBoost, CatBoost, decision tree, ExtraTrees, GBDT, KNN, LightGBM, random forest, SVM, and XGBoost. Each candidate base learner uses the features in the core prediction feature set as input variables and the label of the corresponding binary classification task as output variables. The first base learner layer also includes a hyperparameter optimization unit. This unit employs a Bayesian optimization method, using the combined score of MCC, F1-score, and PR-AUC as the objective function to search for and optimize the hyperparameters of each candidate base learner. The formula for calculating the combined score is as follows: Overall score = 0.4 × MCC + 0.3 × F1-score + 0.3 × PR-AUC.

7. The system for predicting the severity of coronary artery lesions in Kawasaki disease according to claim 1, characterized in that, The interpretability analysis module is implemented based on the SHAP method and includes: The SHAP value calculation unit is used to input the trained Stacking ensemble model, training set background data, and core prediction features into the SHAP interpreter, and calculate the SHAP value of each feature for the prediction result in each sample. The expression for the SHAP value is: ； in, The model represents the sample of children. The predicted output value, This represents the baseline prediction of the model. Indicates the first The contribution value of each feature to the prediction result. Indicates the feature number; The global feature importance analysis unit is used to average the absolute values ​​of each feature in all samples of the training set or validation set to obtain the overall importance of each feature, and output the feature importance ranking in descending order of overall importance; The individualized interpretation unit is used to calculate the SHAP value of each core feature for a single child to be predicted and sort them according to the size of the contribution value. It outputs the main factors that drive the prediction of the child to the corresponding severity and the direction of their influence. The visualization output unit is used to generate feature importance bar charts, SHAP waterfall charts, and SHAP force charts.

8. The system for predicting the severity of coronary artery lesions in Kawasaki disease according to claim 1, characterized in that, The web application deployment module is built on the Streamlit framework and includes: The input interface unit is used to provide input interfaces for each feature in the core prediction feature set. The input interface supports binarized selection input of IVIG resistance, binarized selection input of gender, and numerical input of platelet count, serum albumin, platelet-lymphocyte ratio, total bilirubin, CALLY index and C-reactive protein. Each numerical input box is set with a reasonable numerical range verification. The background prediction unit is used to embed the Stacking integrated learning prediction module and the interpretability analysis module into the web application backend. In response to the prediction trigger operation after the input is completed, it calls the Stacking integrated learning prediction module to perform prediction calculation and calls the interpretability analysis module to generate corresponding visual explanation information. The results display unit is used to simultaneously display the four-level severity prediction probability distribution, the final severity prediction category, the SHAP waterfall plot, and the feature contribution ranking. The results export unit supports the export of prediction results and interpretability analysis reports for inclusion in the patient's medical record.

Images