Autologous arteriovenous fistula patient pta intervention need prediction system and method

Abstract

The present application relates to a kind of autologous arteriovenous fistula patient PTA intervention demand prediction system.The present application also relates to a kind of autologous arteriovenous fistula patient PTA intervention demand prediction method, including data acquisition and candidate variable arrangement process, data preprocessing process, feature screening process, multi-model construction process, model evaluation and optimal model selection process, interpretable nomogram construction process, risk score scale construction and risk stratification threshold setting process, individualized risk prediction process.The present application solves the problems that the prediction target is disconnected with the clinical core decision, the variable dimension is missing, which leads to insufficient prediction efficiency, the performance evaluation is one-sided, which leads to difficult clinical transformation, the machine learning black box characteristic leads to low clinical acceptance and the algorithm professional threshold is high, accurately anchors intervention opportunity, and the model realizes completely interpretable and carries out bedside rapid risk stratification management.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical data analysis and clinical decision support technology, specifically to a system and method for predicting the PTA intervention needs of patients with autogenous arteriovenous fistulas. Background Technology

[0002] In the management of vascular access for maintenance hemodialysis in patients with end-stage renal disease, arteriovenous fistula (AVF) failure is a serious complication, and percutaneous transluminal angioplasty (PTA) is the first-line intervention. However, accurately determining the timing of intervention remains a clinical challenge. Among existing technologies, the closest to this invention are two types of machine learning prediction models focusing on AVF-related complications. One is the AVF stenosis prediction model constructed by Peng Shu et al. in 2025 (Shu P, et al. Eur J Med Res 30, 217(2025)), which incorporates seven basic clinical variables: number of surgeries, prothrombin time activity, lymphocyte count, duration of fistula use, triglycerides, vitamin B12, and C-reactive protein. It constructs seven types of machine learning models, including XGBoost, with a validation set AUC of 0.829. The second model is the arteriovenous fistula thrombosis prediction model constructed by Jinping Ying et al. in 2025 (Ying J, et al. Renal Failure 47, 2477832). This model incorporates variables such as high-sensitivity C-reactive protein (hs-CRP), catheter history, hemodialysis duration, stenosis of the arteriovenous fistula itself, and non-high-density lipoprotein cholesterol. It was constructed using a logistic regression algorithm, and the validation set AUC was 0.826. Both models only used discrimination (AUC value) as the core performance evaluation criterion, without conducting systematic validation of calibration and clinical applicability, and neither included ultrasound-related detection parameters.

[0003] The existing technology has the following significant drawbacks: 1. The research objectives are out of touch with core clinical needs, and no predictive tool has been developed for the "need for PTA intervention after AVF failure".

[0004] 2. The variable dimensions are incomplete, failing to include ultrasound hemodynamic and morphological parameters and systemic inflammation-derived indicators, resulting in limited predictive efficacy; the performance evaluation is one-sided, using only AUC as the evaluation standard, lacking systematic validation for calibration and clinical applicability.

[0005] 3. The models lack interpretability, are mostly black box models, lack quantified weights, lack visual nodal plots, and have not been constructed with scoring scales, making them difficult to implement in clinical practice. Summary of the Invention

[0006] This invention addresses the aforementioned problems by providing a PTA intervention demand prediction system and method for patients with autogenous arteriovenous fistulas. Its purpose is to resolve issues such as the disconnect between prediction targets and core clinical decisions, insufficient predictive efficacy due to missing variable dimensions, difficulties in clinical translation due to one-sided performance evaluation, low clinical acceptance due to the black-box nature of machine learning, and high professional barriers in algorithms. It overcomes the shortcomings of existing AVF prediction models, including prediction targets deviating from actual clinical PTA intervention decision-making needs, missing core ultrasound and inflammation variables, single performance evaluation indicators, uninterpretable black-box models, lack of quantitative evaluation tools, and difficulty in promotion at the grassroots level. This allows for precise targeting of intervention timing, fully interpretable models, and rapid bedside risk stratification management.

[0007] To solve the above problems, the technical solution provided by the present invention is as follows: A system for predicting the need for PTA intervention in patients with autogenous arteriovenous fistulas includes: The data acquisition and candidate variable preparation module is used to extract raw clinical data of hemodialysis patients from electronic medical record systems, laboratory information systems and medical image storage and transmission systems, and to construct a set of candidate predictive variables that includes raw variables and derived variables. The data preprocessing module is used to automatically remove highly missing variables according to a preset missing rate threshold, and fill in the remaining missing data using a chain equation multiple imputation method. The complete dataset is divided into training set and validation set through stratified random sampling. The feature selection module is used to determine the optimal penalty parameters on the training set by using the objective function of the logistic regression model with minimum absolute shrinkage and selection operator, combined with folded cross-validation, and to select the core predictor variable set from the candidate predictor variable set. The multi-model construction and hyperparameter configuration module is used to simultaneously construct binary classification prediction models on the training set based on the core predictive variable set, and determine the key hyperparameters of each binary classification prediction model through grid search combined with the folded cross-validation. The optimal model evaluation and selection module is used to systematically evaluate the binary prediction models on the validation set using the area under the receiver operating characteristic curve (AUC) evaluation model, the mean absolute error (MAE) evaluation model, and the decision curve analysis evaluation model, and determine the optimal prediction model among the binary prediction models. It also determines the optimal classification cutoff threshold based on the Youden exponent maximization method. The interpretable nodal plot transformation module is used to transform the parameters of the optimal prediction model into a visual nodal plot score system based on the optimal prediction model through a general score mapping formula, forward mapping of continuous variables, backward mapping of continuous variables, and binary variable mapping. The module also constructs the mapping relationship between the total score of the nodal plot and the predicted probability of the visual nodal plot score system through the Sigmoid function. The risk scoring scale discretization and control threshold setting module is used to discretize the continuous score system of the nomogram into integer scores. Based on the sample distribution characteristics of the training set, the optimal binary classification segmentation threshold of 37 points is determined based on the Youden exponent maximization method, and patients are divided into two levels: low risk (<37 points) and high risk (≥37 points). The individualized risk prediction and control signal output module is used to calculate the total score of the scale based on the measured values ​​of the core predictive variables of the patient to be assessed, and to calculate the individual PTA intervention prediction probability of the patient through the optimal prediction model. Then, it matches the corresponding risk level and outputs the corresponding management recommendations.

[0008] A method for predicting the PTA intervention needs of patients with autogenous arteriovenous fistulas, comprising the following steps: S100. Data Acquisition and Candidate Variable Preparation Process: Obtain a structured hemodialysis patient dataset, construct a multi-dimensional candidate predictive variable set, and calculate the derived variables; S200. Data preprocessing process: missing data filtering and imputation are performed on the candidate predictor variable set, and then the processed dataset is divided into the training set and the validation set through the stratified random sampling. S300. Feature selection process: Perform feature sparsity estimation using the objective function of the logistic regression model with embedded soft threshold operator to extract the core predictive variable set; S400. Multi-model construction process: Train the binary classification prediction model in parallel on the same feature space and the training set, and determine the key hyperparameters; S500. Model Evaluation and Optimal Model Selection Process: The binary prediction models are quantitatively evaluated on the validation set in multiple dimensions using the area under the receiver operating characteristic curve (AUC) evaluation model, the mean absolute error (MAE) evaluation model, and the decision curve analysis evaluation model. The optimal prediction model is determined and the best classification cutoff threshold is extracted. S600. Interpretable Nomogram Construction Process: Convert the parameters of the optimal prediction model into a continuous score system of the nomogram, and establish a mapping relationship between the total score of the nomogram and the intervention probability based on the Sigmoid function; S700. Risk scoring scale construction and risk stratification threshold setting process: Discretize and round the scores of the continuous score system of the nodal graph, and establish multi-level risk stratification standards; S800. Personalized Risk Prediction Process: Input the measured values ​​of the core predictive variables of the patient to be assessed, automatically trigger the calculation of the total score of the scale and the predicted probability of the individual patient's PTA intervention, determine and output the personalized risk assessment results and the clinical management recommendations.

[0009] Preferably, step S100 specifically includes the following sub-steps: S110. Establish inclusion and exclusion criteria for study subjects: Search clinical databases and extract records of maintenance hemodialysis patients who meet all of the following inclusion criteria and do not trigger any exclusion criteria: The inclusion criteria were: age ≥ 18 years; using autogenous arteriovenous fistula (AVF) as the vascular access for maintenance hemodialysis; and having a complete follow-up medical record, which included demographic information, dialysis records, ultrasound imaging examinations, and laboratory test results. The exclusion criteria are: patients using artificial vascular grafts (AVG) or central venous catheters as dialysis access; patients lost to follow-up or who died during the follow-up period; and patients whose missing data rate for key variables exceeds a preset threshold. S120. Construct a candidate predictor variable system: Collect multi-dimensional candidate predictor variables from the cases screened by S110, and construct a set of multi-dimensional candidate predictor variables including demographic characteristics, dialysis-related information, vital signs, ultrasound imaging indicators, and laboratory test indicators. S130. Calculate the derived variables: extract specific measurement indicators from the multi-dimensional candidate predictive variable set, and calculate the resistance index and the systemic immune inflammation index; The resistance index is expressed by the following formula: Where: PSV is used to characterize the peak systolic velocity, which is the maximum blood flow velocity during the cardiac cycle, in m / s; EDV is the end-diastolic velocity, which is the minimum blood flow velocity at the end of the cardiac cycle, in m / s; The systemic immune inflammatory index is expressed as follows: Wherein: N is used to characterize neutrophil count; P is used to characterize platelet count; L is used to characterize lymphocyte count.

[0010] Preferably, step S200 specifically includes the following sub-steps: S210. Calculate the missing rate and remove the highly missing variables: For each variable in the multi-dimensional candidate predictor variable set, calculate the missing rate, and remove variables whose missing rate is greater than or equal to the missing rate threshold, wherein the missing rate threshold is 30%; the missing rate is expressed by the following formula: in: Used to characterize the Each variable takes the value of the number of missing samples; Used to characterize the total sample size; S220. Processing remaining missing values ​​through multiple imputation: For variables that are retained after removal and satisfy the missing rate <30%, the conditional distribution model is established using the chain equation multiple imputation method to perform iterative imputation. The number of imputations is set to m=5, and 5 independent complete datasets with multiple imputation are generated. S230. Classify the complete multi-imputation dataset: Divide the complete multi-imputation dataset into the training set and the validation set in a 7:3 ratio using a stratified random sampling method.

[0011] Preferably, step S300 specifically includes the following sub-steps: S310. Fitting the objective function of the minimum absolute shrinkage and selection operator: On the training set, construct the objective function of the logistic regression model of the minimum absolute shrinkage and selection operator containing the L1 norm penalty term, expressed as follows: in: Used to characterize the size of the training set; Used to characterize the observation outcome of the i-th sample; Used to characterize the The predicted outcome for each sample is: The probability of; Used to characterize the intercept term; Used to characterize the Regression coefficients of candidate predictor variables; Used to characterize the L1 norm penalty parameter, ; Used to characterize the The sample at the th The values ​​that can be taken on each variable; Used to characterize negative log-likelihood, and used to measure the goodness of fit of the model to the training data; Used to characterize the L1 norm penalty term, which is used to impose a penalty on the sum of the absolute values ​​of the regression coefficients; S320. Soft Threshold Sparsity Estimation: The regression coefficients of the candidate predictor variables are sparsified using the soft threshold operator, which is expressed by the following formula: in: Initial values ​​of coefficients used to characterize ordinary least squares estimation; Used to characterize the sign function; when the initial value of the coefficient of the ordinary least squares estimate is ≤λ, the regression coefficient is directly compressed to zero and the variable is removed from the model; when the initial value of the coefficient of the ordinary least squares estimate is >λ, the regression coefficient is retained and contracted towards zero; S330. Cross-validation parameter tuning: Using the folded cross-validation method with K=5, calculate the binary cross-entropy loss of the L1 norm penalty parameter, and determine the optimal penalty parameter according to the 1se criterion. The optimal penalty parameter is the maximum value that satisfies the requirement that the cross-validation error is within one standard deviation of the minimum error, expressed by the following formula: Wherein: CV is used to characterize the cross-validation error; Used for characterization The standard deviation of the cross-validation error is described above; The penalty parameter used to characterize the minimum cross-validation error; S340. Output the core predictor variable set: On the training set, fit the minimum absolute shrinkage and selection operator logistic regression model with the optimal penalty parameter, and output the core predictor variable set consisting of nine independent core predictor variables, expressed as follows: Where: X RI Used to characterize the resistance index; X Num Used to characterize the number of narrow points; X TypeIV Used to characterize type IV stenosis indicator variables; X BAF Used to characterize brachial artery flow; X SII Used to characterize the systemic immune inflammation index; X DT Used to characterize dialysis duration; X PTAH Used to characterize historical PTA indicators; X HR Used to characterize heart rate; X IT Indicator variable used to characterize intimal thickening.

[0012] Preferably, step S400 specifically includes the following sub-steps: S410. Parallel Construction of Binary Classification Models: Based on the core predictor variable set, six different binary classification prediction models are simultaneously fitted on the same training set, including: Model 1: Logistic Regression Model. This model uses the sigmoid function to map linear combinations of features to probability outputs and employs L2 regularization to control model complexity. The predicted probability of this logistic regression model is expressed by the following formula: in: Used to characterize the sigmoid function; Model 2: Random Forest model, which integrates the prediction results of B=500 decision trees using the Bootstrap meta-algorithm. The sampling ratio of the feature subset of the random forest is set to the square root of the total number of features. The prediction probability of the random forest model is expressed by the following formula: Where: B is used to represent the number of decision trees, B=500; Used to characterize the The predicted probability of category 1 output by each decision tree; Model 3: Support Vector Machine model, which uses a radial basis function kernel to map low-dimensional features to a high-dimensional feature space, and then outputs the predicted probabilities scaled by Platt; the decision function is expressed as follows: in: Used to characterize the support vector index set; Used to characterize Lagrange multipliers; The class labels used to characterize the training samples correspond , correspond ; Used to characterize bias terms; Used to characterize radial basis kernel functions; Used to characterize the Euclidean distance between training samples and input samples; Used to characterize the kernel width parameter; Model 4: Extreme Gradient Boosting Model. Iteratively constructs a weak learner decision tree using a forward-step additive approach. Each tree fits the negative gradient of the current model to gradually reduce the prediction error. The prediction probability of the extreme gradient boosting model is expressed by the following formula: in: Used to characterize The cumulative output of the additive model after rounds of iteration; The sigmoid function is used to characterize the sigmoid function; T is used to characterize the number of decision trees; γ is used to characterize the learning rate. Model 5: Lightweight Gradient Boosting Machine model. Based on the histogram algorithm for discretizing feature intervals, it employs a leaf-wise tree growth strategy to find the optimal split point, significantly improving training efficiency and memory utilization. The histogram algorithm discretizes continuous features into k intervals (default k=255), statistically analyzing gradient information in each interval to quickly determine the optimal split point. The leaf-wise strategy selects the leaf with the largest split gain among all current leaves for splitting, rather than expanding layer by layer (level-wise), thus achieving equivalent model complexity with less tree depth and effectively controlling overfitting while maintaining prediction accuracy. Model Six: Naive Bayes Model. Based on Bayes' theorem and the assumption of conditional independence of features, it uses the Gaussian distribution assumption for continuous variables to estimate the posterior probability density. The predicted probability of the Naive Bayes model is expressed by the following formula: in: Used to characterize categories The prior probability; Used to characterize in category Next The conditional probability of each feature; S420. Grid Search Hyperparameter Configuration: For the six models of the binary classification prediction model, exhaustive combinations are performed within the preset candidate value range, and the classification performance verified by the fold cross-validation method with K=5 is used to determine the final configuration.

[0013] Preferably, step S500 specifically includes the following sub-steps: S510. Quantitative Calculation of AUROC Discrimination: The true positive rate of each model of the binary classification prediction model at different classification thresholds is integrally expressed with respect to the false positive rate using the area under the receiver operating characteristic curve (AUC) assessment model, as follows: in: Used to characterize the The model uses the area under the receiver operating characteristic curve (AUC), which is dimensionless and takes values ​​of [value missing]. , When the value is 0.5, there is no distinguishing ability. Perfectly distinguishable when =1; Used to characterize the Each model at the threshold The true positive rate; Used to characterize the Each model at the threshold The false positive rate; Used to characterize the number of true positive cases; Used to characterize the number of false negatives; Used to characterize the number of false positives; Used to characterize the number of true negative cases; Used to characterize the probability of a classification threshold; S520. Quantitative Calculation of Calibration: The mean absolute error between the average predicted probability of each model in the binary classification prediction model and the actual observation frequency is calculated using the mean absolute error evaluation model, expressed by the following formula: in: Used to characterize the The mean absolute error of each model is dimensionless; G is used to characterize the number of calibration groups. Used to characterize the The model in the first The average predicted probability of the group; Used to characterize the The actual observation frequency of the group; S530. Quantitative Calculation of Clinical Practicality: The clinical practicality of each model of the binary prediction model is evaluated using the decision curve analysis and evaluation model, and the net benefit at each decision threshold is calculated and expressed by the following formula: in: Used to characterize the Each model at the threshold The net gain below is dimensionless; Used to characterize the Each model at the threshold The number of true positive cases; Used to characterize the Each model at the threshold The number of false positives; Used to characterize the size of the validation set; Used to characterize the probability of clinical decision thresholds; S540. Calculate the optimal classification cutoff threshold: The optimal classification cutoff threshold for each model of the binary classification prediction model is determined using the Youden index; the Youden index is expressed by the following formula: in: Used to characterize the The Yoden index of the model is dimensionless. Used to characterize the true negative rate; Used to characterize the optimal classification cutoff threshold; S550. Optimal Model Selection Decision: Among all candidate binary classification prediction models, the optimal prediction model is selected based on the following three quantitative logical judgment criteria, and its final performance index is locked: Validation set robustness judgment logic: Calculate the absolute value of the difference between the area under the receiver operating characteristic curve (AUC) of each binary classification prediction model on the training set and the area under the AUC of the receiver operating characteristic curve (AUC) on the validation set, and select the binary classification prediction model with the smallest absolute value of the difference; Optimal calibration judgment logic: Using the mean absolute error evaluation model, calculate the mean absolute error of each binary classification prediction model on the validation set, and select the binary classification prediction model with the lowest mean absolute error. Clinical applicability is highlighted by the following judgment logic: using the decision curve analysis and evaluation model, the net benefit value of the binary prediction model under each clinical decision threshold probability is calculated. Within the preset clinically commonly used threshold probability range, the net benefit value is selected to be higher than both the extreme intervention strategy line and the extreme non-intervention strategy line, and the net benefit value is higher than other binary prediction models. Based on the comprehensive comparison results of the quantitative logic, the logistic regression model is finally determined as the optimal prediction model, and the area under the receiver operating characteristic curve, the mean absolute error, and the optimal classification cutoff threshold corresponding to the Youden exponent maximization method are locked on the validation set.

[0014] Preferably, step S600 specifically includes the following sub-steps: S610. General Score Mapping: Define the total scale range of the nodal plot, and construct the general score mapping formula based on the regression coefficients of each variable in the optimal prediction model, expressed as follows: in: Used to characterize the The scores of each variable in the nomogram; Used to characterize the total scale range of the nomogram; Used to characterize the maximum absolute value of all regression coefficients; Used to characterize the Measured values ​​of each variable; Used to characterize reference values, taking positive variables. or inverse variable ; Used to characterize the Normalization factor for the range of values ​​of each variable; S620. Execution of branch for differential mapping of three types of variables: Performing the forward mapping of the continuous variables is expressed as follows: ; Performing the inverse mapping of the continuous variables is expressed as follows: ; The binary variable mapping is performed as follows: ; S630. Construct the mapping relationship between the total score of the nodal plot and the intervention probability, expressed as follows: Wherein: TP is used to represent the total score of the nodal chart, and its value ranges from 0 to 100. , The dimensionless parameter is used to characterize the mapping from the total score of the nomogram to the intervention probability. Used to characterize the predictive probability of PTA intervention.

[0015] Preferably, step S700 specifically includes the following sub-steps: S710. Discretized Scoring Algorithm The continuous scores obtained by mapping the continuous score system of the nodal plot in step S600 are discretized into integer scores according to the parameter values ​​using the corresponding discretization algorithm. The discretization algorithms include the forward discretization algorithm for continuous variables, the backward discretization algorithm for continuous variables, and the discretization algorithm for binary variables.

[0016] The continuous variable forward discretization algorithm is applicable to X. RI X Num X SII X DT Expressed as follows: The continuous variable inverse discretization algorithm is applicable to X. BAF X HR Expressed as follows: The binary variable discretization algorithm described above is applicable to X. TypeIV X PTAH X IT Expressed as follows: in: Used to characterize the The scores of each variable in the rating scale are non-negative integers. Used to characterize the The highest score of each variable; round is used to characterize the rounding function; S720. Quantitative division of weight levels: Based on the standardized weight coefficients, the core variables are divided into three levels: the first weight level includes resistance index, number of stenosis and brachial artery flow; the second weight level includes systemic immune inflammation index, dialysis duration, heart rate and type IV stenosis; the third weight level includes past PTA history and intimal thickening. S730. Risk stratification threshold calculation: Based on the sample distribution characteristics of the training set, the optimal binary classification threshold of 37 points is determined by the Youden exponent maximization method, and patients are divided into two levels: low risk (<37 points) and high risk (≥37 points).

[0017] Preferably, step S800 specifically includes the following sub-steps: S810. Variable Collection and Quality Control: Collect the core predictive variable set and execute the following collection quality control rules: the dialysis duration is double-checked against the previous PTA history; the heart rate is limited to measurement before dialysis treatment and before the patient starts dialysis; the brachial artery flow, the resistance index, the number of stenosis, the type IV stenosis indicator variable, and the intimal thickening indicator variable are obtained by physicians with fixed qualifications; the systemic immune inflammation index is limited to blood routine results within two weeks. S820. Calculation of the total score of the rating scale: Substitute the collected measured values ​​into the corresponding discretization formula in step S710, calculate the scores of each item, and sum them to generate the patient's total score of the rating scale, expressed as follows: Wherein: Score is the patient's total score on the rating scale, which is a non-negative integer ranging from 0 to 100.

[0018] S830. Calculation of individual PTA intervention probability: Substituting the measured values ​​of the nine core predictive variables into the optimal predictive model, the predicted probability of individual patient PTA intervention is calculated as follows: in: This is used to characterize the predicted probability of PTA intervention for the individual patient; the closer it is to 1, the more necessary PTA intervention is. S840. Automatic risk level determination: The total score of the rating scale is compared with the binary classification threshold and the risk is automatically classified into low-risk or high-risk levels; S850. Based on the risk level determination result of step S840, output the corresponding management strategy: if classified as low risk level, output the routine follow-up plan, with ultrasound re-examination every 3 months; if classified as high risk level, output the active assessment plan, with immediate comprehensive assessment, arrangement of PTA or surgical intervention, and multidisciplinary consultation.

[0019] Compared with existing technologies, the invention has the following advantages: 1. Because this invention collects candidate variables from five dimensions—demographic characteristics, dialysis information, vital signs, ultrasound parameters, and laboratory indicators—and incorporates nine core predictive indicators after LASSO regression screening, it overcomes the limitations of traditional methods that rely on a single indicator and lack core ultrasound and inflammation variables. The model can achieve a validation set discrimination of up to 0.8806, an ultra-low mean absolute error of 0.067, and a wider range of clinical applicability net benefits.

[0020] 2. Because this invention selects the fully interpretable logistic regression as the optimal prediction model through multi-model system evaluation and transforms it into a visual nodal graph, it converts complex mathematical parameters into an intuitive graphical scale score system. This allows clinicians to make bedside decisions completely transparently and traceably without needing to understand the underlying algorithm, greatly improving the trust in the decision-making process.

[0021] 3. Because this invention constructs a discrete scoring scale of 0-100 points based on nomogram weights and the corresponding binary risk stratification and differentiated clinical management strategies, it can not only achieve individualized precision management and optimize the allocation of medical resources, but also be widely applied in hemodialysis centers of primary hospitals without computer assistance. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall research process in a specific embodiment of the present invention; Figure 2 This is a schematic diagram of LASSO regression screening predictor in a specific embodiment of the present invention; Figure 3-1 This is a schematic diagram of the calibration curve of the LightGBM training set in a specific embodiment of the present invention; Figure 3-2 This is a schematic diagram of the calibration curve of the NB training set in a specific embodiment of the present invention; Figure 3-3 This is a schematic diagram of the calibration curve of the XGBOOST training set in a specific embodiment of the present invention; Figure 3-4 This is a schematic diagram of the calibration curve of the logistic regression training set in a specific embodiment of the present invention; Figure 3-5 This is a schematic diagram of the calibration curve of the random forest training set in a specific embodiment of the present invention; Figure 3-6 This is a schematic diagram of the calibration curve of the support vector machine training set in a specific embodiment of the present invention; Figure 4-1 This is a schematic diagram of the decision curves of the LightGBM validation set in a specific embodiment of the present invention; Figure 4-2 This is a schematic diagram of the decision curve of the NB verification set in a specific embodiment of the present invention; Figure 4-3 A schematic diagram of the decision curves of the XGBOOST validation set in a specific embodiment of the present invention; Figure 4-4 This is a schematic diagram of the decision curves for the logistic regression validation set in a specific embodiment of the present invention. Figure 4-5 This is a schematic diagram of the decision curves of the random forest validation set in a specific embodiment of the present invention; Figure 4-6This is a schematic diagram of the decision curve of the support vector machine validation set in a specific embodiment of the present invention; Figure 5-1 This is a schematic diagram of the calibration curves for the LightGBM verification set in a specific embodiment of the present invention; Figure 5-2 This is a schematic diagram of the calibration curve of the NB verification set in a specific embodiment of the present invention; Figure 5-3 This is a schematic diagram of the calibration curves of the XGBOOST verification set in a specific embodiment of the present invention; Figure 5-4 This is a schematic diagram of the calibration curve of the logistic regression validation set in a specific embodiment of the present invention; Figure 5-5 This is a schematic diagram of the calibration curve of the random forest validation set in a specific embodiment of the present invention; Figure 5-6 This is a schematic diagram of the calibration curve of the support vector machine validation set in a specific embodiment of the present invention; Figure 6-1 This is a schematic diagram of the decision curve of the LightGBM training set in a specific embodiment of the present invention; Figure 6-2 This is a schematic diagram of the decision curve of the NB training set in a specific embodiment of the present invention; Figure 6-3 A schematic diagram of the decision curve of the XGBOOST training set in a specific embodiment of the present invention; Figure 6-4 This is a schematic diagram of the decision curve of the logistic regression training set in a specific embodiment of the present invention. Figure 6-5 This is a schematic diagram of the decision curves of the random forest training set in a specific embodiment of the present invention; Figure 6-6 This is a schematic diagram of the decision curves of the support vector machine training set in a specific embodiment of the present invention; Figure 7-1 This is a schematic diagram of the ROC curve of the LightGBM validation set in a specific embodiment of the present invention; Figure 7-2 This is a schematic diagram of the ROC curve of the NB validation set in a specific embodiment of the present invention; Figure 7-3 This is a schematic diagram of the ROC curve of the XGBOOST verification set in a specific embodiment of the present invention; Figure 7-4 This is a schematic diagram of the logistic regression validation set ROC curve in a specific embodiment of the present invention; Figure 7-5 This is a schematic diagram of the ROC curve of the validation set of a random forest according to a specific embodiment of the present invention; Figure 7-6 This is a schematic diagram of the ROC curve of the support vector machine validation set in a specific embodiment of the present invention; Figure 8-1 This is a schematic diagram of the ROC curve of the LightGBM training set in a specific embodiment of the present invention; Figure 8-2 This is a schematic diagram of the ROC curve of the NB training set in a specific embodiment of the present invention; Figure 8-3 This is a schematic diagram of the ROC curve of the XGBOOST training set in a specific embodiment of the present invention; Figure 8-4 This is a schematic diagram of the ROC curve of the logistic regression training set in a specific embodiment of the present invention; Figure 8-5 This is a schematic diagram of the ROC curve of the random forest training set in a specific embodiment of the present invention; Figure 8-6 This is a schematic diagram of the ROC curve of the support vector machine training set in a specific embodiment of the present invention; Figure 9 This is a schematic diagram of a nomogram showing the Points scale axis, total-to-score axis, and PTA intervention risk, which includes nine core predictive variables, according to a specific embodiment of the present invention. Detailed Implementation

[0023] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading the present invention, any modifications of the present invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0024] It should be noted beforehand that, for the convenience of readers, the applicant has listed the terms and their English equivalents and abbreviations involved in the following text here, in the format of "Chinese (full English name, English abbreviation)".

[0025] Percutaneous transluminal angioplasty (PTA); arteriovenous fistula (AVF); Area Under the Receiver Operating Characteristic Curve (AUROC); Least Absolute Shrinkage and Selection Operator (LASSO); Multivariate Imputation by Chained Equations (MICE); Dialysis Time (DT); Heart Rate (HR); History of Previous PTA (PTA_History); Brachial Artery Flow (BA_Flow); Resistance Index (RI); Number of Stenosis (Stenosis Numbers); Type IV stenosis (Type IV); Intimal Thickening (IT); Systemic Immune Inflammatory Index (SII). Immune-Inflammation Index (SII); Mean Absolute Error (MAE); Random Forest (RF); Support Vector Machine (SVM); Logistic Regression (LR); Extreme Gradient Boosting (XGBoost); Light Gradient Boosting Machine (LightGBM); Naive Bayes (NB); Type IV stenosis (Type IV); Decision Curve Analysis (DCA); Electronic Medical Record (EMR); Laboratory Information System (LIS)Picture Archiving and Communication Systems (PACS); Peak Systolic Velocity (PSV); End Diastolic Velocity (EDV); Resistive Index (RI); Neutrophil Count (N); Lymphocyte Count (L); Platelet Count (P); High-sensitivity C-reactive Protein (hs-CRP); C-reactive Protein (CRP); White Blood Cell (WBC); Hemoglobin (Hb); Hematocrit (HCT); Body Mass Index (BMI); Total Cholesterol (TC); Triglycerides (TG); Low-Density Lipoprotein Cholesterol (LDL-C) Cholesterol (LDL); Parathyroid Hormone (PTH); Peritoneal Dialysis (PD); Nomogram; Points; Binomial Deviance; Coefficients; Bootstrap (B=500 replicates); Mean Absolute Error; Predicted Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve; Sample Size 290 / 124 (n=290 / n=124); Cost-Benefit Ratio; High Risk Threshold; All Intervention; None Intervention; Specificity; Sensitivity; 95% Confidence Interval (95% CI); Scores, Points; Total Score Points);Disease risk; cut-off value; Youden index; nomogram; risk score scale; individualized risk prediction. Figure 1 Random Forest (RF); Support Vector Machine (SVM); Logistic Regression; Extreme Gradient Boosting Tree (XGBoost); Lightweight Gradient Boosting Tree (LightGBM); Naive Bayes (NB); Figure 2 Binomial deviation; regression coefficients. Figure 3-1 Predicted Probability; Lightweight Gradient Boosting Tree (LightGBM); Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve; Figure 3-2 Predicted Probability; Naive Bayes (NB); Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve; Figure 3-3 Predicted Probability; XGBoost; Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve. Figure 3-4 : Predicted Probability; Logistic Regression; Observed probability; Ideal curve; Apparent curve; Bias-corrected curve; Figure 3-5Predicted Probability; Random Forest (RF); Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve; Figure 3-6 Predicted Probability; Support Vector Machine (SVM); Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve; Figure 4-1 Lightweight Gradient Boosting Tree (LightGBM); All Intervention; None Intervention; Cost-Benefit Ratio; High Risk Threshold; Standardized Net Benefit. Figure 4-2 Naive Bayes (NB); All intervention; None intervention; Cost-Benefit Ratio; High Risk Threshold; Standardized Net Benefit. Figure 4-3 Extreme gradient boosting tree (XGBoost); All intervention; None intervention; Cost-benefit ratio; High risk threshold; Standardized net benefit. Figure 4-4 Logistic Regression; All Intervention; None Intervention; Cost-Benefit Ratio; High Risk Threshold; Standardized Net Benefit. Figure 4-5Random Forest (RF); All intervention; None intervention; Cost-Benefit Ratio; High Risk Threshold; Standardized Net Benefit. Figure 4-6 Support Vector Machine (SVM); All intervention; None intervention; Cost-Benefit Ratio; High Risk Threshold; Standardized Net Benefit. Figure 5-1 Lightweight Gradient Boosting Tree (LightGBM); Predicted Probability; Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve; Figure 5-2 Predicted Probability; Naive Bayes (NB); Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve. Figure 5-3 Predicted Probability; XGBoost; Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve. Figure 5-4 Predicted Probability; Logistic Regression; Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve; Figure 5-5Predicted Probability; Random Forest (RF); Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve; Figure 5-6 Predicted Probability; Support Vector Machine (SVM); Observed Probability; Ideal Curve; Apparent Curve; Bias-corrected Curve; Figure 6-1 Lightweight Gradient Boosting Tree (LightGBM); All Intervention; None Intervention; Cost-Benefit Ratio; High Risk Threshold; Standardized Net Benefit. Figure 6-2 Naive Bayes (NB); All intervention; None intervention; Cost-Benefit Ratio; High Risk Threshold; Standardized Net Benefit. Figure 6-3 : Extreme gradient boosting tree (XGBoost); All intervention; None intervention; Cost-benefit ratio; High-risk threshold; Standardized net benefit. Figure 6-4 Logistic Regression; All Intervention; None Intervention; Cost-Benefit Ratio; High Risk Threshold; Standardized Net Benefit. Figure 6-5Random Forest (RF); All intervention; None intervention; Cost-Benefit Ratio; High Risk Threshold; Standardized Net Benefit. Figure 6-6 Support Vector Machine (SVM); All intervention; None intervention; Cost-Benefit Ratio; High Risk Threshold; Standardized Net Benefit. Figures 7-1 to 8-6 Specificity; Sensitivity.

[0026] This invention application claims a system for predicting the need for PTA intervention in patients with autogenous arteriovenous fistulas, comprising: The data acquisition and candidate variable preparation module is used to extract raw clinical data of hemodialysis patients from electronic medical record systems, laboratory information systems and medical image storage and transmission systems, and to construct a set of candidate predictive variables that includes raw variables and derived variables. The data preprocessing module is used to automatically remove highly missing variables according to a preset missing rate threshold, and fill in the remaining missing data using a chain equation multiple imputation method. The complete dataset is divided into training set and validation set through stratified random sampling. The feature selection module is used to determine the optimal penalty parameters by using the objective function of the logistic regression model with minimum absolute shrinkage and selection operator on the training set, combined with folded cross-validation, and to select the core predictor variable set from the candidate predictor variable set. The multi-model construction and hyperparameter configuration module is used to simultaneously construct binary classification prediction models on the training set based on the core predictor variable set, and determine the key hyperparameters of each binary classification prediction model through grid search combined with folded cross-validation. The optimal model evaluation and selection module is used to systematically evaluate the model on the validation set using the area under the receiver operating characteristic curve (AUC) evaluation model, the mean absolute error (MAE) evaluation model, and the decision curve analysis evaluation model, and to determine the optimal prediction model among the binary prediction models. It also determines the optimal classification cutoff threshold based on the Youden exponent maximization method. The interpretable nodal plot transformation module is used to transform the parameters of the optimal prediction model into a visual nodal plot score system based on the optimal prediction model through general score mapping formulas, forward mapping of continuous variables, backward mapping of continuous variables, and mapping of binary variables. It also constructs the mapping relationship between the total score of the nodal plot and the predicted probability of the visual nodal plot score system through the Sigmoid function. The risk scoring scale discretization and control threshold setting module is used to discretize the continuous score system of the nomogram into integer scores. Based on the sample distribution characteristics of the training set, the optimal binary classification threshold of 37 points is determined by the Youden exponent maximization method, and patients are divided into two levels: low risk (<37 points) and high risk (≥37 points). The individualized risk prediction and control signal output module is used to calculate the total score of the scale based on the measured values ​​of the core predictive variables of the patient to be evaluated, and to calculate the individual PTA intervention prediction probability of the patient through the optimal prediction model. Then, it matches the corresponding risk level and outputs the corresponding management recommendations.

[0027] This application also claims a method for predicting the need for PTA intervention in patients with autogenous arteriovenous fistulas, such as... Figure 1 As shown, the process includes the following: S100. Data Acquisition and Candidate Variable Preparation Process: Obtain a structured dataset of hemodialysis patients, construct a multi-dimensional set of candidate predictive variables, and calculate derived variables; S200. Data preprocessing process: Perform missing data filtering and imputation on the candidate predictor variable set, and then divide the processed dataset into training set and validation set through stratified random sampling; S300. Feature selection process: Perform feature sparsity estimation using the objective function of the logistic regression model with embedded soft threshold operator and logistic selection operator to extract the core predictive variable set; S400. Multi-model construction process: Train binary classification prediction models in parallel on the same feature space and training set, and determine key hyperparameters; S500. Model Evaluation and Optimal Model Selection Process: The model is evaluated by using the area under the receiver operating characteristic curve, the mean absolute error, and the decision curve analysis. The model is quantitatively evaluated on the validation set in multiple dimensions to determine the optimal prediction model and extract the best classification cutoff threshold. S600. Interpretable Nomogram Construction Process: Convert the parameters of the optimal prediction model into a continuous score system for the nomogram, and establish a mapping relationship between the total score of the nomogram and the intervention probability based on the Sigmoid function; S700. Risk scoring scale construction and risk stratification threshold setting process: Discretize and round the scores of the continuous score system of the nodal chart, and establish multi-level risk stratification standards; S800. Personalized Risk Prediction Process: Input the measured values ​​of the core predictive variables of the patient to be assessed, automatically trigger the calculation of the total scale score and the individual patient's PTA intervention prediction probability, determine and output the personalized risk assessment results and clinical management recommendations.

[0028] It should be noted that step S100 aims to screen maintenance hemodialysis patients who meet the study requirements from the clinical database and establish an initial study cohort. Specifically, it includes the following sub-steps: S110. Establish inclusion and exclusion criteria for study subjects: Search clinical databases and extract records of maintenance hemodialysis patients who meet all of the following inclusion criteria and do not trigger any exclusion criteria: Inclusion criteria were: age ≥ 18 years; use of autogenous arteriovenous fistula (AVF) as vascular access for maintenance hemodialysis; and possession of complete follow-up medical records, including demographic information, dialysis records, ultrasound imaging examinations, and laboratory test results. Exclusion criteria were: patients using AVG grafts or central venous catheters as dialysis access; patients lost to follow-up or who died during the follow-up period; and patients whose missing data rate for key variables exceeded the preset threshold. S120. Construct a candidate predictive variable system: Collect multi-dimensional candidate predictive variables from cases screened through S110, and construct a multi-dimensional candidate predictive variable set including demographic characteristics, dialysis-related information, vital signs, ultrasound imaging indicators, and laboratory test indicators. S130. Calculate derived variables: Extract specific measurement indicators from the multi-dimensional candidate predictor variable set and calculate the resistance index and the systemic immune inflammation index; the resistance index and the systemic immune inflammation index are two important derived variables. The resistance index is expressed by the following formula (1): (1) Where: PSV is used to characterize the peak systolic velocity, which is the maximum blood flow velocity during the cardiac cycle, in m / s; EDV is the end-diastolic velocity, which is the minimum blood flow velocity at the end of the cardiac cycle, measured in m / s; the resistance index X RI The value ∈[0,1] indicates that the greater the distal vascular resistance and the worse the blood perfusion, and is a key ultrasound indicator for assessing the patency of arteriovenous fistulas.

[0029] Systemic immune inflammatory indexes are expressed as follows (2): (2) Wherein: N represents neutrophil count; P represents platelet count; L represents lymphocyte count. X SII XSII comprehensively reflects the balance between the body's neutrophil-mediated pro-inflammatory state and lymphocyte-mediated immune state; a higher value indicates a stronger systemic inflammatory response. In this embodiment, XSII is calculated using the most recent complete blood count result before enrollment; if an acute infection occurred within the past week, the first test result after infection control was used.

[0030] It should be further noted that, in this embodiment, the data were collected from a retrospective cohort of maintenance hemodialysis patients at the Hemodialysis Center of Wuhan Central Hospital from January 2021 to December 2023. Clinical data were extracted jointly from the hospital's Electronic Medical Record (EMR), Laboratory Information System (LIS), and Picture Archiving and Communication System (PACS). A total of 414 eligible cases were included according to the above inclusion and exclusion criteria. Data collection was approved by the hospital's ethics committee.

[0031] It should be further explained that those meeting the inclusion criteria... For example, patient data was collected and processed systematically. There are 10 candidate predictor variables, which constitute the candidate predictor variable set. These 56 variables are categorized into the following five groups based on their clinical attributes and data collection sources: The first category is demographic characteristics: There are approximately eight variables, including age, sex, body mass index, marital status, smoking history, and drinking history, which are used to describe the patient's basic sociological and physiological background.

[0032] The second category, dialysis-related information, comprises approximately 10 variables, including dialysis duration XDT (unit: days, representing the total duration from the start of hemodialysis to this assessment), duration of use of autogenous arteriovenous fistula, and history of previous percutaneous endovascular angioplasty XDT. PTAH (Binary variable, X) PTAH =1 indicates that the patient has previously received PTA intervention, dialysis frequency, etc.

[0033] The third category is vital signs: approximately six variables, including heart rate X. HR (Unit: beats / min), systolic blood pressure, diastolic blood pressure, etc.

[0034] The fourth category is ultrasound imaging parameters: This includes approximately 18 variables, including brachial artery flow X. BAF (Unit: mL / min), Drag Index X RI (Dimensionless), Narrow Quantity X Num (Unit: number) Type IV stenosis X TypeIV (Binary variable, X) TypeIV=1 indicates diffuse stenosis involving the entire anastomosis and outflow tract; X indicates intimal thickening. IT (Binary variable, X) IT =1 indicates the presence of endometrial thickening, etc.

[0035] The fifth category comprises laboratory test indicators: approximately 14 variables, including hemoglobin, white blood cell count, neutrophil count (N), lymphocyte count, platelet count, parathyroid hormone, total cholesterol, triglycerides, and systemic immune inflammatory index (X). SII wait.

[0036] In this embodiment, all ultrasound imaging parameters were acquired using a GE Healthcare LOGIQ E11 ultrasound diagnostic instrument equipped with a 7-15 MHz linear array probe and a preset vascular / small organ program. The brachial artery flow rate X... BAF With resistance index X RI The measurement location was uniformly set at the widest point of the blood flow jet in the mid-segment of the brachial artery in the upper arm (5-10 cm above the elbow crease), with a sampling volume of 2-3 mm. The angle between the sound beam and the blood flow was controlled within 60°, and the average value of three consecutive cardiac cycles was taken. All laboratory test parameters were obtained from tests conducted within one week prior to the ultrasound examination.

[0037] It should be further explained that the completion of this step is marked by: completing the screening of 414 study subjects, establishing a set of candidate predictor variables containing P=56 candidate predictor variables, completing the calculation of all derived variables, and forming a structured dataset that can be used for subsequent analysis.

[0038] It should be noted that step S200 specifically includes the following sub-steps: S210. Calculate the missing rate and remove highly missing variables: For each variable in the multidimensional candidate predictor set, calculate the missing rate and remove variables whose missing rate is greater than or equal to the missing rate threshold, where the missing rate threshold is 30%; the missing rate is expressed by the following formula (3): (3) in: Used to characterize the Each variable takes the value of the number of missing samples; Used to characterize the total sample size; S220. Handling remaining missing values ​​through multiple imputation: For variables that are retained after removal and have a missing rate of <30%, a conditional distribution model is established using the chain equation multiple imputation method to perform iterative imputation. The number of imputations is set to m=5, generating 5 independent complete datasets with multiple imputation. S230. Classify the complete multi-imputation dataset: Divide the complete multi-imputation dataset into training and validation sets using stratified random sampling at a ratio of 7:3.

[0039] It should be further noted that when the variable missing rate exceeds 30%, the remaining sample size is insufficient to guarantee the robustness of subsequent statistical inferences and may introduce systematic bias.

[0040] In this embodiment, variables removed due to a missing rate ≥30% include some nutritional indicators and inflammatory markers such as serum albumin, transferrin saturation, and C-reactive protein (CRP). Although these variables have clinical predictive value, their detection rate was insufficient and the information incomplete during retrospective data collection, thus failing to meet the inclusion criteria. After this screening step, the remaining variables proceed to the next processing stage.

[0041] It should be further noted that a small amount of missing data (R) still exists among the variables retained after step S210. miss,j For variables less than 30%, multivariate imputation by chained equations (MICE) is used for imputation. MICE generates m complete datasets by building a conditional distribution model for each incomplete variable and iteratively drawing random samples from each conditional distribution. In this step, the number of imputations is set to m=5, resulting in 5 independent multivariate imputation complete datasets.

[0042] For each variable containing missing values, MICE establishes a conditional regression model with that variable as the dependent variable and the remaining variables as independent variables (continuous variables use predicted mean matching, and binary variables use logistic regression), iteratively updating the missing value estimate until convergence. The five complete datasets generated from five independent imputations are then combined and analyzed in subsequent steps to reflect the impact of imputation uncertainty on the model results.

[0043] It should be further explained that the stratification in step S230 is based on the category distribution of the outcome variable Y, to ensure that the sample proportions of Y=1 (requiring PTA intervention) and Y=0 (not requiring intervention) in the two subsets remain consistent.

[0044] In this embodiment, the partitioning result is: training set sample size N train =290, validation set sample size N val =124, satisfying N train +N val =N total =414. Training set D train Used for model fitting and parameter estimation, validation set D valThis data is used solely for model performance evaluation and does not participate in any training process. The training and validation sets are compared based on demographic characteristics, clinical history, dialysis parameters, ultrasound parameters, and S... II There were no statistically significant differences in key variables ( ). p (>0.05), indicating good sample balance and inter-group comparability.

[0045] The completion of this step is marked by: removing missing variables, generating m=5 complete datasets with multiple imputation, and establishing the training set D using stratified random sampling at a ratio of 7:3. train (N) train =290) and validation set D val (N) val =124), ensuring that the outcome variable Y is distributed consistently in both subsets.

[0046] It should be noted that, as Figure 2 As shown, step S300 specifically includes the following sub-steps: S310. Fitting the objective function of the minimum absolute contraction and selection operator: On the training set, construct the objective function of the logistic regression model with the minimum absolute contraction and selection operator containing the L1 norm penalty term, expressed as follows (4): (4) in: Used to characterize the size of the training set; Used to characterize the observation outcome of the i-th sample; Used to characterize the The predicted outcome for each sample is: The probability of; Used to characterize the intercept term; Used to characterize the Regression coefficients of candidate predictor variables; Used to characterize the L1 norm penalty parameter, ; Used to characterize the The sample at the th The values ​​that can be taken on each variable; Used to characterize negative log-likelihood, and used to measure the goodness of fit of the model to the training data; Used to characterize the L1 norm penalty term, which is used to impose a penalty on the sum of the absolute values ​​of the regression coefficients; S320. Soft Threshold Sparsity Estimation: The regression coefficients of candidate predictors are sparsified by using a soft threshold operator, which is expressed by the following formula (5): (5) in: Initial values ​​of coefficients used to characterize ordinary least squares estimation; Used to characterize the sign function; when the initial value of the coefficient of ordinary least squares estimation is ≤λ, the regression coefficient is directly compressed to zero and the variable is removed from the model; when the initial value of the coefficient of ordinary least squares estimation is >λ, the regression coefficient is retained and contracted towards zero; thus, L1 penalty has the dual functions of variable selection and coefficient contraction.

[0047] S330. Cross-validation parameter tuning: Using the K=5 folded cross-validation method, calculate the binary cross-entropy loss of the L1 norm penalty parameter, and determine the optimal penalty parameter according to the 1se criterion. The optimal penalty parameter is the maximum value that satisfies the requirement that the cross-validation error is within one standard deviation of the minimum error, and is expressed by the following formula (6): (6) Wherein: CV is used to characterize the cross-validation error; Used for characterization The standard deviation of the cross-validation error; The penalty parameter used to characterize the minimum cross-validation error; S340. Output the core predictor set: On the training set, fit the minimum absolute shrinkage and selection operator logistic regression model with the optimal penalty parameter, and output the core predictor set consisting of nine independent core predictors, expressed as follows (7): (7) Where: X RI Used to characterize the resistance index; X Num Used to characterize the number of narrow points; X TypeIV Used to characterize type IV stenosis indicator variables; X BAF Used to characterize brachial artery flow; X SII Used to characterize the systemic immune inflammation index; X DT Used to characterize dialysis duration; X PTAH Used to characterize historical PTA indicators; X HR Used to characterize heart rate; X IT Indicator variable used to characterize intimal thickening.

[0048] It should be further explained that step S310 aims to select a core set of variables from 56 candidate predictor variables that have independent predictive power for the need for postoperative percutaneous intravascular angioplasty intervention. Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression is used for feature selection. LASSO introduces an L1 norm penalty term into the log-likelihood function of logistic regression, simultaneously achieving variable selection and coefficient shrinkage, effectively addressing variable redundancy and multicollinearity issues in high-dimensional data. train=290 is the number of training set samples.

[0049] It should be further noted that the value of the penalty parameter λ in step S330 directly determines the number of variables retained in the model and the degree of coefficient shrinkage, and needs to be determined through a data-driven approach. This step uses K=5-fold cross-validation to fine-tune λ.

[0050] The specific process is as follows: The training set D... train (N) train =290) is randomly divided into K=5 disjoint subsets (each subset contains approximately 58 samples). For a given λ, the LASSO model is fitted sequentially using the k-th subset as the validation fold and the union of the remaining four subsets as the training fold, and the binary cross-entropy loss is calculated on the validation fold. This process is repeated 5 times and the average is taken to obtain the cross-validation error under that λ. λ is chosen as follows. 1se The principle behind the optimal penalty parameter is to select the largest λ value to obtain the sparsest model, while ensuring that the model's predictive performance does not decrease significantly (within one standard deviation of the minimum bias), thereby reducing the risk of overfitting and improving the interpretability of the model.

[0051] In this embodiment, LASSO regression is implemented using the glmnet package in R language. The elastic network mixture parameter α is set to 1 (corresponding to pure L1 regularization), and the loss function family is set to "binomial" (binary logistic regression). The training set feature matrix is ​​standardized and then input into the model. Five-fold cross-validation is used, and the mean squared deviation of cross-validation (Deviance) is used as the evaluation metric. The cross-validation curve is plotted on the logarithmic scale sequence of λ. λ is selected. 1se As the optimal penalty parameter, 9 core predictor variables were retained from 56 candidate variables.

[0052] It should be further noted that the set of variables removed in step S340 is V. removed Satisfying V removed =V candidates \V core And V removed =47.

[0053] The 47 removed variables can be categorized into the following four types: The first type, those with high missing rates: although satisfying R... miss,j <30% but still relatively high missing rate, incomplete information limits predictive contribution, such as some nutritional indicators and inflammatory markers; the second category, strongly collinear: redundant variables highly correlated with retained variables, such as anastomotic diameter, proximal artery diameter, etc. RI X BAFHighly correlated (r>0.6), various types of narrow classification and X TypeIV X Num Information overlap; the third category, weak associations: variables whose marginal association strength with variable Y is insufficient to reach the LASSO retention threshold, such as gender, BMI, marital status, total cholesterol, triglycerides, etc.; the fourth category, variables covered by derived variables: original variables whose predictive information has been summarized by derived composite indicators, such as white blood cell count, neutrophil count, etc., have been covered by derived variables. SII Comprehensive coverage.

[0054] The nine core predictor variables retained after LASSO screening and their regression coefficients are shown in Table 1 below: Table 1. Nine core predictor variables retained after LASSO screening and their regression coefficients. The completion of this step is indicated by: based on the training set D train Complete LASSO logistic regression fitting, using 5-fold cross-validation and λ... 1se The criteria determine the optimal penalty parameters, and from 56 candidate variables, 9 core predictor variables are selected to establish the core predictor variable set V. core .

[0055] It should be noted that the nine core predictor variables V obtained based on the S300 steps are... core ={RI,Num,TypeIV,BAF,SII,DT,PTAH,HR,IT}, this step constructs 6 different binary classification prediction models (denoted as the k-th classification model, k=1,2,…,6) to compare the performance of different algorithms in this prediction task. All models are trained on the same set D. train (N) train Fitting was performed on a dataset of 290 examples (p=9 features). Hyperparameter tuning was determined using a grid search combined with 5-fold cross-validation. The validation set D... valid (N) valid =124 examples) do not participate in any hyperparameter tuning process. For example Figure 3-1 , 3-2 As shown in 3-3, 3-4, 3-5, 3-6, 4-1, 4-2, 4-3, 4-4, 4-5, 4-6, 5-1, 5-2, 5-3, 5-4, 5-5, 5-6, 6-1, 6-2, 6-3, 6-4, 6-5, 6-6, 7-1, 7-2, 7-3, 7-4, 7-5, 7-6, 8-1, 8-2, 8-3, 8-4, 8-5, 8-6, step S400 specifically includes the following sub-steps: S410. Parallel Construction of Binary Classification Models: Based on the core predictor variable set, six different binary classification prediction models are simultaneously fitted on the same training set, including: Model 1: Logistic regression model, which uses the sigmoid function to map the linear combination of features into probability output, and uses L2 regularization to control the model complexity. The predicted probability of the logistic regression model is expressed by the following formula (8): (8) in: Used to characterize the sigmoid function; β0 is the intercept term; β j For the j-th core predictor variable x j The regression coefficients, j=1,2,…,9, β j ∈ℝ. Logistic regression employs L2 regularization (Ridge Regularization), and the regularization parameter C is determined through grid search combined with 5-fold cross-validation.

[0056] The advantages of logistic regression lie in its simple model structure and the direct interpretability of its parameters (the exponential form of the regression coefficients exp(β)). j Corresponding ratio OR j ), suitable as a baseline model.

[0057] Implementation data: The AUROC of logistic regression on the training set was 0.9147 (95% CI: 0.8841-0.9453), with a sensitivity of 0.880, specificity of 0.807, and MAE of 0.061; on the validation set, the AUROC was 0.8806 (95% CI: 0.8179-0.9432), with a sensitivity of 0.800, specificity of 0.763, and MAE of 0.067. The AUROC decay from the training set to the validation set was only 0.0341, demonstrating excellent generalization ability.

[0058] Model 2: Random Forest model, which integrates the prediction results of B=500 decision trees using the Bootstrap meta-algorithm. Each tree is trained independently on randomly selected sample subsets and feature subsets, and the final output is the mean of the prediction probabilities of all trees. The feature subset sampling ratio of the random forest is set to the square root of the total number of features (…). This increases the diversity among trees and reduces the risk of overfitting. The prediction probability of the random forest model is expressed by the following equation (9): (9) Where: B is used to represent the number of decision trees, B=500; Used to characterize the The predicted probability of category 1 is output by each decision tree.

[0059] Implementation data: The AUROC of Random Forest on the training set is 0.8672 (95% CI: 0.8268-0.9077), with a sensitivity of 0.787, a specificity of 0.807, and a MAE of 0.089; on the validation set, the AUROC is 0.8611 (95% CI: 0.7937-0.9286), with a sensitivity of 0.723, a specificity of 0.814, and a MAE of 0.095. The AUROC decreases by 0.0061 from the training set to the validation set.

[0060] Model 3: Support Vector Machine (SVM) model. SVM maps low-dimensional features to a high-dimensional feature space using a radial basis function kernel, and then searches for the optimal hyperplane that maximizes the class margin in this space. The predicted probabilities are then output after Platt scaling; the decision function is expressed as follows (10): (10) in: Used to characterize the support vector index set; Used to characterize Lagrange multipliers; The class labels used to characterize the training samples correspond , correspond ; Used to characterize bias terms; Used to characterize radial basis kernel functions; Used to characterize the Euclidean distance between training samples and input samples; The kernel width parameter is used to characterize the kernel function and controls its local influence range. The model output is mapped to probability values ​​using Platt scaling, which means that the SVM's decision function output is calibrated to a probability estimate through an additional logistic regression.

[0061] Implementation data: The AUROC of the support vector machine on the training set is 0.9307 (95% CI: 0.9023-0.9590), with a sensitivity of 0.833, a specificity of 0.900, and a MAE of 0.072; on the validation set, the AUROC is 0.8761 (95% CI: 0.8148-0.9375), with a sensitivity of 0.677, a specificity of 0.847, and a MAE of 0.088. The AUROC decreases by 0.0546 from the training set to the validation set.

[0062] Model 4: Extreme Gradient Boosting Model. This model iteratively constructs weak learner decision trees using a forward-step additive approach. Each tree fits the negative gradient of the current model to gradually reduce prediction error. Let the total number of iterations be T, the output scaling factor (learning rate) of each tree be γ, and the output of the t-th tree be... The prediction probability of the extreme gradient boosting model is expressed by the following equation (11): (11) in: Used to characterize The cumulative output of the additive model after rounds of iteration; The sigmoid function is used to represent the number of decision trees; γ is used to represent the learning rate; T=300; γ=0.05, controlling the contribution of each tree to the overall model to prevent overfitting. Generalization ability is further improved through regularization terms (L1 / L2 penalties on the number of leaf nodes and leaf weights) and column / row sampling.

[0063] Implementation data: XGBoost achieved an AUROC of 0.9609 (95% CI: 0.9427-0.9791) on the training set, with a sensitivity of 0.827, specificity of 0.957, and MAE of 0.048. However, on the validation set, the AUROC dropped to 0.8606 (95% CI: 0.7951-0.9262), with a sensitivity of 0.723, specificity of 0.831, and MAE of 0.112. The AUROC decreased by as much as 0.1003 from the training set to the validation set, and the calibration MAE deteriorated from 0.048 to 0.112, indicating significant overfitting.

[0064] Model 5: Lightweight Gradient Boosting Machine (LightGBM) is based on the gradient boosting framework, but it significantly improves training efficiency and memory utilization through the histogram-based algorithm and the leaf-wise tree growth strategy.

[0065] The histogram algorithm discretizes continuous features into k intervals (default k=255) and statistically analyzes gradient information in each interval to quickly determine the optimal split point. The leaf-wise strategy selects the leaf with the largest split gain among all current leaves for splitting each time, rather than expanding layer by layer (level-wise), thus achieving the same model complexity with less tree depth and effectively controlling overfitting while maintaining prediction accuracy.

[0066] Implementation data: LightGBM achieved an AUROC of 0.9685 (95% CI: 0.9524-0.9846) on the training set, with a sensitivity of 0.860, specificity of 0.950, and MAE of 0.042, making it the best performing model on the training set among the six. However, on the validation set, the AUROC dropped to 0.8639 (95% CI: 0.7992-0.9285), with a sensitivity of 0.692, specificity of 0.847, and MAE of 0.105. The AUROC decreased significantly from the training set to the validation set, dropping by as much as 0.1046, and the calibration MAE deteriorated from 0.042 to 0.105, indicating a more severe overfitting problem compared to XGBoost.

[0067] Model 6: Naive Bayes model, based on Bayes' theorem and the assumption of conditional independence of features, uses the Gaussian distribution assumption for continuous variables to estimate the posterior probability density; the predicted probability of the Naive Bayes model is expressed by the following formula (12): (12) in: Used to characterize categories The prior probability; Used to characterize in category Next The conditional probabilities of each feature are estimated using the Gaussian Naive Bayes assumption for density estimation of continuous variables. Although the feature independence assumption is often not strictly true in practice, Naive Bayes is frequently used as a simple and effective benchmark model due to its computational efficiency and robustness to high-dimensional data.

[0068] Implementation data: The AUROC of Naive Bayes on the training set is 0.9030 (95% CI: 0.8695-0.9366), with a sensitivity of 0.833, specificity of 0.829, and MAE of 0.078; on the validation set, the AUROC is 0.8699 (95% CI: 0.8067-0.9330), with a sensitivity of 0.769, specificity of 0.847, and MAE of 0.091. The AUROC decreases by 0.0331 from the training set to the validation set.

[0069] S420. Grid Search Hyperparameter Configuration: For six binary classification prediction models, exhaustive combinations are performed within a preset candidate value range. The classification performance is verified using a K=5 fold cross-validation method to determine the final configuration. See Table 2 below: Table 2. Configuration of Key Hyperparameters of the Model During the grid search process, the key hyperparameters of each model are exhaustively combined within the preset candidate value range. The classification performance of 5-fold cross-validation (AUROC as the optimization objective) is used as the evaluation criterion, and the parameter combination that optimizes the validation performance is selected as the final configuration.

[0070] The completion of this step is indicated by: based on the core predictor variable set V core The nine variables in the training set D train The construction and hyperparameter tuning of six classification models have been completed, and each model is now capable of outputting the probability of PTA intervention requirement 6 months postoperatively. The ability.

[0071] It should be noted that after the model training is completed, the six candidate models (logistic regression, random forest, support vector machine, extreme gradient boosting, lightweight gradient boosting, and Naive Bayes) are validated on the validation set D. valid (N) valid Three systematic evaluations were conducted on 124 cases to determine the optimal model based on a comprehensive assessment of discrimination, calibration, and clinical applicability. Step S500 specifically includes the following sub-steps: S510. Quantitative calculation of AUROC discrimination score: The ability of the model to distinguish PTA intervention needs is evaluated by integrating the true positive rate of each binary prediction model with respect to the false positive rate at different classification thresholds using the Area Under the Receiver Operating Characteristic Curve (AUROC) model, as expressed by the following formula (13): (13) in: Used to characterize the The model uses the area under the receiver operating characteristic curve (AUC), which is dimensionless and takes values ​​of [value missing]. , When the value is 0.5, there is no distinguishing ability. Perfectly distinguishable when =1; Used to characterize the Each model at the threshold The true positive rate; Used to characterize the Each model at the threshold The false positive rate; Used to characterize the number of true positive cases; Used to characterize the number of false negatives; Used to characterize the number of false positives; Used to characterize the number of true negative cases; Used to characterize the probability of a classification threshold; S520. Quantitative Calculation of Calibration: The mean absolute error between the average predicted probability of each model in the binary classification prediction model and the actual observation frequency is calculated using the Mean Absolute Error (MAE) evaluation model, and expressed by the following formula (14): (14) in: Used to characterize the The mean absolute error of each model is dimensionless; G is used to characterize the number of calibration groups, G=10 (decimal grouping). Used to characterize the The model in the first The average predicted probability of the group; Used to characterize the The actual observation frequency of the group; S530. Quantitative Calculation of Clinical Practicality: The clinical practicality of each model of the binary prediction model is evaluated using decision curve analysis. The net benefit at each decision threshold is calculated and expressed as follows (15): (15) in: Used to characterize the Each model at the threshold The net gain below is dimensionless. Superior to all non-intervention strategies; Used to characterize the Each model at the threshold The number of true positive cases; Used to characterize the Each model at the threshold The number of false positives; Used to characterize the size of the validation set. =124; Used to characterize the probability of clinical decision thresholds; if a model is within a relatively wide range... Within the range If the values ​​are all higher than other models and the two extreme reference lines ("all intervention line" and "no intervention line"), then the model has better clinical applicability.

[0072] S540. Calculate the optimal classification cutoff threshold: Use the Youden index to determine the optimal classification cutoff threshold for each model of the binary classification prediction model; the Youden index is expressed by the following formula (16): (16) in: Used to characterize the The Yoden index of the model is dimensionless. Used to characterize the true negative rate; Used to characterize the optimal classification cutoff threshold; S550. Optimal Model Selection Decision: Among all candidate binary classification prediction models, the optimal prediction model is selected based on the following five quantitative logical judgment criteria, and its final performance index is determined: The validation set robustness assessment logic is as follows: The absolute value of the difference between the area under the receiver operating characteristic (AUC) curve on the training set and the AUC curve on the validation set for each binary classification prediction model is calculated, and the binary classification prediction model with the smallest absolute difference is selected. The LR validation set AUC is 0.8806 (95% CI: 0.8179-0.9432), with the highest AUC for the six classes. The training set AUC is 0.9147, with a decay of only 0.0341, indicating excellent generalization ability. Although XGBoost and LightGBM achieve AUCs of 0.9609 and 0.9685 respectively on the training set, they drop to 0.8606 and 0.8639 respectively on the validation set, with AUC decay rates as high as 0.1003 and 0.1046 respectively, indicating significant overfitting.

[0073] The optimal calibration judgment logic is as follows: using the mean absolute error (MAE) evaluation model, the mean absolute error of each binary classification prediction model on the validation set is calculated, and the binary classification prediction model with the lowest mean absolute error is selected; the MAE of the LR validation set is 0.067, and the six models have the lowest MAE (RF: 0.095, SVM: 0.088, XGBoost: 0.112, LightGBM: 0.105, NB: 0.091), indicating the highest consistency between the predicted probability and the actual observed probability.

[0074] Clinical practicality is a key criterion: The decision curve analysis is used to evaluate the model, calculating the net benefit of the binary prediction model at various clinical decision threshold probabilities. Within a pre-defined range of commonly used clinical threshold probabilities, the model is selected based on its net benefit, which is simultaneously higher than both the extreme intervention and non-intervention strategy lines, and also higher than other binary prediction models. Decision curve analysis shows that LR (Regression Result) at the threshold probability... Net benefit within the clinically commonly used interval [0.10, 0.60] Both strategies outperformed the two extreme strategies of no intervention and full intervention, and were also superior to other comparative models.

[0075] Full interpretability: regression coefficient β j It directly reflects the direction and magnitude of each variable's contribution to the log odds, and can present the prediction logic completely transparently without relying on posterior interpretation methods such as SHAP.

[0076] Direct conversion of nodal plots: LR linear predictor It can be directly converted into nomograms and rating scales, achieving a seamless transformation from algorithm to bedside clinical tools and supporting rapid bedside calculations.

[0077] The final performance metric of the optimal logistic regression model on the validation set is: AUROC = 0.8806 (95% CI: 0.8179–0.9432), and the optimal cutoff threshold is c. =0.505 (Youden exponent maximization method), sensitivity = 0.800, specificity = 0.763, calibration MAE = 0.067.

[0078] The completion of this step is indicated by: LR being determined as the optimal model, and the validation set AUROC = 0.8806, MAE = 0.067, and c... =0.505.

[0079] Based on the comprehensive comparison results of quantitative logic, the logistic regression model was finally determined as the optimal prediction model. On the validation set, the area under the receiver operating characteristic curve, mean absolute error, and the optimal classification cutoff threshold corresponding to the Youden exponent maximization method were locked for this model.

[0080] The complete performance comparison of the six candidate models on the training and validation sets is shown in the table below: Table 3. Performance comparison of the six types of models on the training set Table 4. Performance comparison of the six models on the validation set It should be noted that in step S600, the regression coefficients of the optimal LR model are... β j This is converted into a visual nomogram, allowing clinicians to intuitively assess individual PTA intervention risk without needing to understand the underlying algorithms. The total scale range of the nomogram is 0% to 100%.

[0081] Specifically, it includes the following sub-steps: S610. General score mapping: Set the total scale range of the nomogram, and construct a general score mapping formula based on the regression coefficients of each variable in the optimal prediction model, expressed as follows (17): (17) in: Used to characterize the The scores of each variable in the nomogram; Used to characterize the total scale range of the nomogram, this embodiment uses 100 points; Used to characterize the maximum absolute value of all regression coefficients; Used to characterize the Measured values ​​of each variable; Used to characterize reference values, taking positive variables. or inverse variable ; Used to characterize the Normalization factor for the range of values ​​of each variable; S620. Execution of branch for differential mapping of three types of variables: Perform a positive mapping of continuous variables (increasing the variable size increases the risk; apply beta). j >0, such as X RI X Num X SII X DT Expressed as follows (18): (18); Perform inverse mapping for continuous variables (increasing the number of variables reduces risk; beta is applicable). j <0, such as X BAF X HR Expressed as follows (19): (19); Perform binary classification variable mapping, applicable to X TypeIV X PTAH X IT It can be expressed as follows (20): (20); S630. Construct the mapping relationship between the total score of the nodal plot and the intervention probability, expressed by the following formula (21): (twenty one) Wherein: TP is used to represent the total score of the nodal chart, and its value ranges from 0 to 100. , The dimensionless parameter is used to characterize the mapping from the total score of the nomogram to the intervention probability. Used to characterize the predictive probability of PTA intervention.

[0082] It should be noted that the variables in step S620 are sorted in descending order of the absolute value of their regression coefficients, as shown in the table below: Table 5. Regression coefficients, OR values, and highest scores of nomograms for core predictor variables sorted by weight hierarchy. In clinical use, the operator positions the measured values ​​of each patient variable on the corresponding scale axis of the nomogram, reads the scores downwards along the vertical lines of each variable's scale, and accumulates them to obtain TP. Then, the TP is positioned on the total score scale line and reads the corresponding P value downwards. PTAThe graphical presentation of nomograms allows clinicians to intuitively assess individual risk without performing logarithmic probability calculations, making them suitable for bedside decision-making scenarios without computer assistance.

[0083] Step S700 transforms the continuous scoring system of the nodal chart into a discrete scoring scale to adapt to the needs of rapid bedside assessment, and sets a two-tier threshold to establish a hierarchical management system. Specifically, it includes the following sub-steps: S710. The continuous scores obtained by mapping the continuous score system of the nodal plot in step S600 are discretized into integer scores according to the parameter values ​​using the corresponding discretization algorithm; the discretization algorithm includes continuous variable forward discretization algorithm, continuous variable backward discretization algorithm, and binary variable discretization algorithm. Forward Discretization Algorithm for Continuous Variables in X RI X Num X SII X DT It can be expressed as follows (22): (twenty two) The continuous variable inverse discretization algorithm is applicable to X BAF X HR It can be expressed as follows (23): (twenty three) The binary classification variable discretization algorithm is suitable for X. TypeIV X PTAH X IT Expressed as follows (24): (twenty four) in: The value of the j-th variable in the rating scale is a non-negative integer. The function used to represent the highest score of the j-th variable is 'j'; 'round' is used to represent the rounding function.

[0084] The highest scores for each variable and the discretization rules are shown in Table 6 below: Table 6. Highest Score for Each Variable and Discretization Rules Continuous variable ratings employ piecewise linear mapping: for positively correlated variables (X... RI X Num X SII X DT The larger the measured value, the higher the corresponding score; for inverse relationship variables (X) BAF X HR The larger the measured value, the lower the corresponding score. Binary variables are assigned values ​​directly based on the state. The maximum total score of the rating scale is given by formula (25): (25) Total score range: 0-100 points.

[0085] S720. Quantitative division of weight levels: Based on standardized weight coefficients, the core variables are divided into three levels: the first weight level includes resistance index, number of stenosis and brachial artery flow; the second weight level includes systemic immune inflammation index, dialysis duration, heart rate and type IV stenosis; the third weight level includes past PTA history and intimal thickening.

[0086] First weighted tier: High-weighted core factors: number of stenosis (maximum 23 points), brachial artery flow (maximum 14 points), and resistance index (maximum 11 points). The combined score of these three factors ranges from 0 to 48 points, accounting for 48% of the scale's full score. They are the core driving factors for the model's predictive effectiveness and the main criteria for risk stratification.

[0087] The second weighted tier consists of medium-weighted risk factors: heart rate (maximum 14 points), systemic immune inflammation index (maximum 10 points), dialysis duration (maximum 9 points), and type IV stenosis (maximum 9 points). The combined score of these four factors ranges from 0 to 42 points, accounting for 42% of the scale's full score, and serves as an important supplement to risk stratification.

[0088] The third weighted level: low-weighted auxiliary factors: history of PTA, endometrial thickening (maximum 5 points each). The combined score of these two factors ranges from 0 to 10 points, accounting for 10% of the total score, and has a limited impact on the overall score.

[0089] S730. Risk stratification threshold calculation: Based on the sample distribution characteristics of the training set, the optimal binary classification threshold of 37 points is determined by the Youden exponent maximization method, and patients are divided into two levels: low risk (<37 points) and high risk (≥37 points).

[0090] Table 7. Risk Level Parameter Table It should be noted that, as Figure 9 As shown, step S800 specifically includes the following sub-steps: S810. Variable Collection and Quality Control: A set of core predictive variables was collected, and the following quality control rules were implemented: Dialysis duration and previous PTA history were double-checked; discrepancies were submitted to a third party for arbitration. Heart rate was limited to measurements before dialysis treatment and before the patient started dialysis to avoid interference from volume changes during dialysis. Brachial artery flow, resistance index, number of stenosis, type IV stenosis indicator variables, and intimal thickening indicator variables were measured by the same physician with over 5 years of experience in vascular ultrasound, using a GE Healthcare LOGIQ E11 ultrasound diagnostic instrument and a 7-15MHz linear array probe. Systemic immune inflammation index was limited to blood routine results within two weeks. The measured values ​​of the nine core predictive variables collected from the patient are shown in Table 8 below. Table 8. Nine core predictive variables for patient data collection S820. Calculation of the total score of the rating scale: Substitute the collected measured values ​​into the corresponding discretization formula in step S710, calculate the scores of each item and sum them to generate the patient's total score of the rating scale, expressed as follows (26): (26) Wherein: Score is the patient's total score on the rating scale, which is a non-negative integer ranging from 0 to 100.

[0091] S830. Calculation of individual PTA intervention probability: Substitute the measured values ​​of the nine core predictive variables into the optimal prediction model to calculate the individual patient PTA intervention prediction probability, expressed as follows (27): (27) Where: P PTA The value is used to characterize the predictive probability of PTA intervention for an individual patient; the closer it is to 1, the more necessary the PTA intervention is. β0 = -2.045513 is the logistic regression intercept term; βj is the regression coefficient of the j-th variable (see Table 5). S840. Automatic Risk Level Determination: The total score of the rating scale is compared with the risk stratification threshold of 37 points to automatically classify the risk into a low-risk or high-risk level; as shown in the following formula (28): Risk=Lif Score<37 Hif Score≥37 (28) In the formula, L represents the low-risk level, corresponding to a score of <37 points; H represents the high-risk level, corresponding to a score of ≥37 points.

[0092] S850. Based on the risk level determination result of step S840, output the corresponding management strategy: R low Output the routine follow-up plan (ultrasound re-examination every 3 months); R highPropose an active assessment plan (immediate comprehensive assessment, arrange PTA or surgical intervention, multidisciplinary consultation).

[0093] Clinical rapid assessment procedure The procedure for rapid clinical assessment based on rating scales is as follows: Step 1: Data Collection: Review patient medical records and extract values ​​for 9 core indicators. DT and X PTAH Obtained from the electronic medical record system; X HR Obtained by bedside measurement before dialysis; X BAF X RI X Num X TypeIV and X IT Obtained from the most recent ultrasound report (all ultrasound indicators mentioned above are routine AVF assessment items); X SII Calculated from the most recent complete blood count result (X SII =N×P / L).

[0094] Step 2: Item-by-item scoring: Referring to the paper scoring scale, map the measured values ​​of each of the 9 indicators to their corresponding scores S. j And record it. Continuous variables are evaluated using piecewise linear interpolation or by taking the score of the nearest anchor point, while binary variables are directly scored according to their state.

[0095] Step 3 Total Score Calculation and Risk Assessment: The scores of the 9 items are added together to obtain the Score. The patient's risk level is determined according to the risk stratification criteria: <37 points is low risk, and ≥37 points is high risk.

[0096] Step 4: Output management recommendations: Output corresponding clinical management recommendations based on the risk level: Low risk, routine follow-up (ultrasound every 3 months); High risk, active intervention (immediate assessment, arrangement of PTA or surgery, multidisciplinary consultation).

[0097] Applicable scenarios: Rapid risk stratification of patients' AVF status during routine ward rounds in hemodialysis centers; regular risk assessment of maintenance hemodialysis patients during outpatient follow-ups; immediate risk quantification of patients suspected of having abnormal AVF function during bedside consultations; screening of high-risk patients by dialysis room nurses before patients start dialysis and notifying physicians for focused assessment; preliminary risk screening in primary care hemodialysis centers when there is no specialist ultrasound physician.

[0098] Typical clinical application cases Case 1: Low-risk patient (58-year-old male, dialysis for 180 days, score 2) Measured values ​​and scores of 9 core variables: The nomogram score is approximately 10.4, corresponding to a predicted probability of approximately 0.5%. The simplified scoring scale score is 2, falling into the low-risk range (<37 points). Management recommendations: Maintain the routine dialysis and follow-up protocol, with an ultrasound examination every 3 months.

[0099] Case 2: High-risk borderline patient (62-year-old female, dialysis for 362 days, score 37) Measured values ​​and scores of 9 core variables: The nomogram score is approximately 34.7, corresponding to a predicted probability of approximately 81%. The simplified scoring scale score is 37, which is within the risk stratification threshold (≥37). Management recommendations: Immediately conduct a comprehensive AVF functional assessment, including angiography to clarify the anatomical details of the stenosis; actively arrange PTA or surgical intervention, prioritizing lesions with significant hemodynamic stenosis; initiate multidisciplinary consultation to develop a comprehensive treatment plan; monitor every 2 weeks until the intervention is completed, and manage at a low-risk level after the intervention.

[0100] Case 3: High-risk patient (55-year-old male, dialysis for 1860 days, score 80) Measured values ​​and scores of 9 core variables: The nomogram score is approximately 75.8, corresponding to a prediction probability of approximately 99.9%. The simplified rating scale score is 80, falling into the high-risk range (≥37 points).

[0101] Management recommendations: Conduct a comprehensive AVF functional assessment immediately, arrange PTA or surgical intervention, initiate multidisciplinary consultation to develop a comprehensive treatment plan, and monitor every 2 weeks until the intervention is completed.

[0102] A comprehensive comparison of three typical cases: Three case studies illustrate the complete clinical decision-making pathway of the predictive system. Low-risk patients exhibit protective characteristics such as short dialysis time, normal brachial artery flow, few stenosis, and no structural abnormalities. Patients at the risk stratification threshold have some overlapping risk factors (increased stenosis, intimal thickening, low flow), and their scores are at the risk stratification threshold. High-risk patients have almost all risk factors pointing to high risk (long dialysis time, low heart rate, history of PTA, low brachial artery flow, high resistance index, multiple stenosis with type IV stenosis and other structural abnormalities, and significantly elevated SII).

[0103] Signs of completion of individualized risk prediction task The prediction task is complete when the patient receives all three of the following outputs: The first item is the predicted probability P of PTA intervention. PTA Quantify the likelihood of PTA intervention; The second risk level assessment result is R. low 、or R high one; The third item, which corresponds to the risk level, provides individualized clinical management recommendations that directly guide subsequent management plans.

[0104] The above three outputs constitute a complete individualized risk prediction report, which can be directly incorporated into the patient's dialysis access management file as a basis for clinical decision support.

[0105] The following is supplementary technical information related to this invention: (This supplementary technical information is helpful for a comprehensive understanding of this patent proposal, covering eight aspects: patient baseline characteristic statistics, multi-model performance comparison, core variable detection specifications, nomogram parameter manual, ultrasound operation specifications, typical clinical cases, technical limitation analysis, and model implementation code. It provides detailed reference for those skilled in the art to repeat the implementation of this invention.) 1. Detailed statistical table of core patient characteristics Complete baseline data of 414 maintenance hemodialysis patients were collected. Continuous variables were described as mean ± standard deviation or median (interquartile range), categorical variables were described as frequency (percentage), and independent samples were used for comparisons between groups. Test, Mann-Whitney Test or chi-square test.

[0106] Appendix 1. Patient Baseline Characteristics (Training Set) vs Validation Set ) The baseline characteristics comparison above showed no statistically significant differences between the training and validation sets in key variables such as demographic characteristics (age, sex, BMI), clinical history (history of diabetes, history of hypertension, previous PTA), dialysis indicators (dialysis duration, AVF usage duration), ultrasound imaging parameters (brachial artery blood flow, resistance index, number of stenosis, type IV stenosis, intimal thickening), and SII (p>0.05). Among all analyzed variables, only alcohol consumption history showed a statistically significant difference between the two groups (p=0.017, 2% in the training set vs. 8% in the validation set). This variable was not one of the nine core predictors included in this model and had a weak direct association with PTA intervention needs. Furthermore, the proportion of alcohol-consuming patients in both groups was low, limiting its impact on model construction and validation. Besides alcohol consumption history, no statistically significant differences were observed between the two groups in non-core indicators such as total cholesterol (p=0.902), aortic calcification (p=0.223), and anastomosis method (p=1.000). The above data fully demonstrates that the training and validation sets have good sample balance and inter-group comparability, laying a reliable statistical foundation for subsequent model construction and validation. This section provides the basis for comprehensive model evaluation and selection. For detailed performance comparison data on the training and validation sets, please refer to Tables 3 and 4 in step S550 of the main text.

[0107] Appendix 2. Model Comprehensive Evaluation and Selection Criteria Optimal Model Selection Analysis: The criteria for choosing LR are as follows: First, LR exhibits the best validation set robustness. While XGBoost and LightGBM have high AUROCs on their training sets (0.9609 and 0.9685, respectively), their validation set AUROCs drop to 0.8606 and 0.8639, with decay rates of 0.1003 and 0.1046, indicating overfitting. LR, on the other hand, has a validation set AUROC of 0.8806, with a decay rate of only 0.0341. Second, LR demonstrates the best calibration with a validation set MAE of 0.067, lower than all other models. Third, LR demonstrates outstanding clinical applicability. DCA shows that LR provides the greatest net benefit within the threshold probability range of 0.10-0.60. Fourth, LR provides fully interpretable regression coefficients that intuitively present the contributions of each variable without requiring a complex computational environment. Fifth, LR can be directly converted into nomograms and rating scales, realizing the transformation from algorithm to bedside tool.

[0108] Appendix 3. Definitions and Testing Standards for 9 Core Predictive Variables Key points for quality control of variable data collection: (1) X DT and X PTAH Two researchers independently verify the electronic medical records; if there are discrepancies, the matter is submitted to a third party for arbitration; (2) X HRMeasurements should be taken uniformly before dialysis treatment and before the patient starts dialysis to avoid interference from volume changes during dialysis; (3) Ultrasound-related variables (X) BAF X RI X Num X TypeIV X IT The procedure should be performed by the same physician with more than 5 years of experience in vascular ultrasound to reduce operator variability. (4) XSII calculation should ensure that the time interval between blood routine test and enrollment does not exceed 2 weeks. If there has been a recent blood transfusion, acute infection or use of drugs that affect blood cell count, it should be noted in the record.

[0109] Operating Procedures for Testing Core Parameters of Ultrasonic Diagnostic Instruments This system has 5 ultrasound parameters (X) BAF X RI X Num X TypeIV X IT All samples were collected using GE Healthcare LOGIQE11.

[0110] Appendix 5. Equipment Configuration Requirements Standard procedure for measuring each parameter (1) Brachial artery flow (X) BAF Measurement: The patient lies supine with the upper limb abducted 30°-45°. The probe is placed on the anteromedial aspect of the upper arm, displaying the brachial artery in the long-axis view. The widest point of the blood flow jet is located in the middle segment of the brachial artery, 5-10 cm above the cubital fossa. After activating color Doppler to confirm blood flow filling, switch to pulsed Doppler mode, sampling volume 2-3 mm, beam-blood flow angle <60° (ideally ≤45°), acquire the spectrum, freeze the image, and take the average blood flow value over 3 consecutive cardiac cycles, in mL / min.

[0111] (2) Resistance Index (X) RI Measurement: with X BAF Data is acquired synchronously at the same measurement point. The equipment calculates automatically. The result is rounded to two decimal places and is the average of three consecutive cardiac cycles.

[0112] (3) Determination and counting of stenotic lesions (X) Num Scan the entire AVF (vascular drainage vein) starting from the anastomosis. A reduction in vessel diameter of >50% in grayscale mode indicates significant morphological stenosis; accelerated blood flow and turbulence on color Doppler suggest stenosis; a local PSV increase of >2 times compared to the adjacent normal segment in PW mode indicates significant hemodynamic stenosis. Count multiple stenosis points within the same vessel segment separately.

[0113] (4) Type IV stenosis (X TypeIVDetermination: Scan the entire length of the AVF (vasodilation reflux tract) starting from the anastomosis to assess the extent of stenosis. Stenosis involving the entire anastomosis and outflow tract (diffuse stenosis) is determined to be type IV stenosis (code 1); otherwise, it is not present (code 0).

[0114] (5) Endometrial thickening (X) IT Judgment: The intima-media thickness (IMT) is measured at the thickest point of the AVF reflux vein cross-section in high-resolution grayscale mode, accurate to 0.1 mm. IMT > 0.6 mm is defined as intimal thickening (code 1), and ≤ 0.6 mm is normal (code 0).

[0115] Key points of quality control: All ultrasound examinations are performed by the same physician with more than 5 years of experience in vascular ultrasound; the equipment is calibrated in a standardized manner before each examination; for continuous variables, the average of 3 consecutive measurements is taken, and if the difference between a measurement and the average is >10%, the measurement is repeated; all images are saved in DICOM format.

[0116] Core code snippet of the model The following code is written in R (v4.3.1), with key dependencies: glmnet (LASSO regression), rms (nomogram), caret (training framework), pROC (ROC analysis), and rmda (DCA).

[0117] LASSO regression variable selection code Code Explanation: The core implementation of the prediction system consists of three modules. Module 1 selects the optimal subset of variables from high-dimensional candidate features through 5-fold cross-validation; Module 2 fits the logistic regression model and provides a complete performance evaluation (AUC, sensitivity, specificity, calibration, DCA); Module 3 transforms the model into a visual noctilinear plot and a corresponding rating scale.

[0118] Data Source Notes: All data are from a retrospective cohort of maintenance hemodialysis patients at the Hemodialysis Center of Wuhan Central Hospital from January 2021 to December 2023, approved by the hospital's ethics committee (WHZXKYL2023-016), with informed consent waived. Ultrasound data were acquired using a GE LOGIQ E11, laboratory data were obtained from the LIS system, and clinical history was obtained from the EMR system. The model was developed using R 4.3.1, with the following main package versions: glmnet 4.1-7, rms 6.7-0, caret 6.0-94, pROC 1.18.4, and rmda 1.6.

[0119] Appendix 6 below is a PTA intervention risk quantification scoring table constructed based on the risk contribution weights of each variable in a nomogram according to a specific embodiment of the present invention: Appendix 6. PTA Intervention Risk Quantification Scoring Table Constructed Based on Risk Contribution Weights of Variables in Nonograph Appendix Table 7 below is a PTA intervention risk stratification table based on the total score of the quantitative scoring table using a dichotomy method, according to a specific embodiment of the present invention: Appendix 7. Risk Stratification Table for PTA Intervention Based on Dichotomy Using Total Score of Quantitative Scoring Scale In the above detailed description, various features are combined together in a single embodiment to simplify this disclosure. This approach to disclosure should not be construed as reflecting an intention that embodiments of the claimed subject matter require more features than are explicitly stated in each claim. Rather, as reflected in the appended claims, the invention is presented with fewer features than all of the features of the single disclosed embodiment. Therefore, the appended claims are hereby explicitly incorporated into the detailed description, wherein each claim stands alone as a preferred embodiment of the invention.

[0120] The disclosed embodiments have been described above to enable any person skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments without departing from the spirit and scope of this disclosure. Therefore, this disclosure is not limited to the embodiments given herein, but is consistent with the broadest scope of the principles and novel features disclosed in this application.

[0121] The foregoing description includes examples of one or more embodiments. It is certainly impossible to describe all possible combinations of components or methods in order to describe the above embodiments, but those skilled in the art will recognize that further combinations and arrangements of the various embodiments are possible. Therefore, the embodiments described herein are intended to cover all such changes, modifications, and variations that fall within the scope of the appended claims. Furthermore, the term "comprising" as used in the specification or claims is interpreted in a manner similar to the term "including," as interpreted when used as a conjunction in the claims. Additionally, the use of any term "or" in the specification of the claims is intended to mean "non-exclusive or."

[0122] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A system for predicting the need for PTA intervention in patients with autogenous arteriovenous fistulas, characterized in that: include: The data acquisition and candidate variable preparation module is used to extract raw clinical data of hemodialysis patients from electronic medical record systems, laboratory information systems and medical image storage and transmission systems, and to construct a set of candidate predictive variables that includes raw variables and derived variables. The data preprocessing module is used to automatically remove highly missing variables according to a preset missing rate threshold, and fill in the remaining missing data using a chain equation multiple imputation method. The complete dataset is divided into training set and validation set through stratified random sampling. The feature selection module is used to determine the optimal penalty parameters on the training set by using the objective function of the logistic regression model with minimum absolute shrinkage and selection operator, combined with folded cross-validation, and to select the core predictor variable set from the candidate predictor variable set. The multi-model construction and hyperparameter configuration module is used to simultaneously construct binary classification prediction models on the training set based on the core predictive variable set, and determine the key hyperparameters of each binary classification prediction model through grid search combined with the folded cross-validation. The optimal model evaluation and selection module is used to systematically evaluate the binary prediction models on the validation set using the area under the receiver operating characteristic curve (AUC) evaluation model, the mean absolute error (MAE) evaluation model, and the decision curve analysis evaluation model, and determine the optimal prediction model among the binary prediction models. It also determines the optimal classification cutoff threshold based on the Youden exponent maximization method. The interpretable nodal plot transformation module is used to transform the parameters of the optimal prediction model into a visual nodal plot score system based on the optimal prediction model through a general score mapping formula, forward mapping of continuous variables, backward mapping of continuous variables, and binary variable mapping. The module also constructs the mapping relationship between the total score of the nodal plot and the predicted probability of the visual nodal plot score system through the Sigmoid function. The risk scoring scale discretization and control threshold setting module is used to discretize the continuous score system of the nomogram into integer scores. Based on the sample distribution characteristics of the training set, the optimal binary classification segmentation threshold of 37 points is determined based on the Youden exponent maximization method, and patients are divided into two levels: low risk (<37 points) and high risk (≥37 points). The individualized risk prediction and control signal output module is used to calculate the total score of the scale based on the measured values ​​of the core predictive variables of the patient to be assessed, and to calculate the individual PTA intervention prediction probability of the patient through the optimal prediction model. Then, it matches the corresponding risk level and outputs the corresponding management recommendations.

2. A method for predicting the PTA intervention needs of patients with autogenous arteriovenous fistulas as described in claim 1, characterized in that: The process includes the following steps: S100. Data Acquisition and Candidate Variable Preparation Process: Obtain a structured hemodialysis patient dataset, construct a multi-dimensional candidate predictive variable set, and calculate the derived variables; S200. Data preprocessing process: missing data filtering and imputation are performed on the candidate predictor variable set, and then the processed dataset is divided into the training set and the validation set through the stratified random sampling. S300. Feature selection process: Perform feature sparsity estimation using the objective function of the logistic regression model with embedded soft threshold operator to extract the core predictive variable set; S400. Multi-model construction process: Train the binary classification prediction model in parallel on the same feature space and the training set, and determine the key hyperparameters; S500. Model Evaluation and Optimal Model Selection Process: The binary prediction models are quantitatively evaluated on the validation set in multiple dimensions using the area under the receiver operating characteristic curve (AUC) evaluation model, the mean absolute error (MAE) evaluation model, and the decision curve analysis evaluation model. The optimal prediction model is determined and the best classification cutoff threshold is extracted. S600. Interpretable Nomogram Construction Process: Convert the parameters of the optimal prediction model into a continuous score system of the nomogram, and establish a mapping relationship between the total score of the nomogram and the intervention probability based on the Sigmoid function; S700. Risk scoring scale construction and risk stratification threshold setting process: Discretize and round the scores of the continuous score system of the nodal graph, and establish multi-level risk stratification standards; S800. Personalized Risk Prediction Process: Input the measured values ​​of the core predictive variables of the patient to be assessed, automatically trigger the calculation of the total score of the scale and the predicted probability of the individual patient's PTA intervention, determine and output the personalized risk assessment results and the clinical management recommendations.

3. The method according to claim 2, characterized in that: Step S100 specifically includes the following sub-steps: S110. Establish inclusion and exclusion criteria for study subjects: Search clinical databases and extract records of maintenance hemodialysis patients who meet all of the following inclusion criteria and do not trigger any exclusion criteria: The inclusion criteria were: age ≥ 18 years; using autogenous arteriovenous fistula (AVF) as the vascular access for maintenance hemodialysis; and having a complete follow-up medical record, which included demographic information, dialysis records, ultrasound imaging examinations, and laboratory test results. The exclusion criteria are: patients using AVG grafts or central venous catheters as dialysis access; patients lost to follow-up or who died during the follow-up period; Patients whose key variable data missing rate exceeds a preset threshold; S120. Construct a candidate predictor variable system: Collect multi-dimensional candidate predictor variables from the cases screened by S110, and construct a set of multi-dimensional candidate predictor variables including demographic characteristics, dialysis-related information, vital signs, ultrasound imaging indicators, and laboratory test indicators. S130. Calculate the derived variables: extract specific measurement indicators from the multi-dimensional candidate predictive variable set, and calculate the resistance index and the systemic immune inflammation index; The resistance index is expressed by the following formula: Where: PSV is used to characterize the peak systolic velocity, which is the maximum blood flow velocity during the cardiac cycle, in m / s; EDV is the end-diastolic velocity, which is the minimum blood flow velocity at the end of the cardiac cycle, in m / s; The systemic immune inflammatory index is expressed as follows: Wherein: N is used to characterize neutrophil count; P is used to characterize platelet count; L is used to characterize lymphocyte count.

4. The method according to claim 2, characterized in that: Step S200 specifically includes the following sub-steps: S210. Calculate the missing rate and remove the highly missing variables: For each variable in the multi-dimensional candidate predictor variable set, calculate the missing rate, and remove variables whose missing rate is greater than or equal to the missing rate threshold, wherein the missing rate threshold is 30%; the missing rate is expressed by the following formula: in: Used to characterize the Each variable takes the value of the number of missing samples; Used to characterize the total sample size; S220. Processing remaining missing values ​​through multiple imputation: For variables that are retained after removal and satisfy the missing rate <30%, the conditional distribution model is established using the chain equation multiple imputation method to perform iterative imputation. The number of imputations is set to m=5, and 5 independent complete datasets with multiple imputation are generated. S230. Classify the complete multi-imputation dataset: Divide the complete multi-imputation dataset into the training set and the validation set in a 7:3 ratio using a stratified random sampling method.

5. The method according to claim 2, characterized in that: Step S300 specifically includes the following sub-steps: S310. Fitting the objective function of the minimum absolute shrinkage and selection operator: On the training set, construct the objective function of the logistic regression model of the minimum absolute shrinkage and selection operator containing the L1 norm penalty term, expressed as follows: in: Used to characterize the size of the training set; Used to characterize the observation outcome of the i-th sample; Used to characterize the The predicted outcome for each sample is: The probability of; Used to characterize the intercept term; Used to characterize the Regression coefficients of candidate predictor variables; Used to characterize the L1 norm penalty parameter, ; Used to characterize the The sample at the th The values ​​that can be taken on each variable; Used to characterize negative log-likelihood, and used to measure the goodness of fit of the model to the training data; Used to characterize the L1 norm penalty term, which is used to impose a penalty on the sum of the absolute values ​​of the regression coefficients; S320. Soft Threshold Sparsity Estimation: The regression coefficients of the candidate predictor variables are sparsified using the soft threshold operator, which is expressed by the following formula: in: Initial values ​​of coefficients used to characterize ordinary least squares estimation; Used to characterize symbolic functions; When the initial value of the coefficient estimated by ordinary least squares is ≤ λ, the regression coefficient is directly compressed to zero and the variable is removed from the model; when the initial value of the coefficient estimated by ordinary least squares is > λ, the regression coefficient is retained and contracted towards zero. S330. Cross-validation parameter tuning: Using the folded cross-validation method with K=5, calculate the binary cross-entropy loss of the L1 norm penalty parameter, and determine the optimal penalty parameter according to the 1se criterion. The optimal penalty parameter is the maximum value that satisfies the requirement that the cross-validation error is within one standard deviation of the minimum error, expressed by the following formula: Wherein: CV is used to characterize the cross-validation error; Used for characterization The standard deviation of the cross-validation error is described above; The penalty parameter used to characterize the minimum cross-validation error; S340. Output the core predictor variable set: On the training set, fit the minimum absolute shrinkage and selection operator logistic regression model with the optimal penalty parameter, and output the core predictor variable set consisting of nine independent core predictor variables, expressed as follows: Where: X RI Used to characterize the resistance index; X Num Used to characterize the number of narrow points; X TypeIV Used to characterize type IV stenosis indicator variables; X BAF Used to characterize brachial artery flow; X SII Used to characterize the systemic immune inflammation index; X DT Used to characterize dialysis duration; X PTAH Used to characterize historical PTA indicators; X HR Used to characterize heart rate; X IT Indicator variable used to characterize intimal thickening.

6. The method according to claim 2, characterized in that: Step S400 specifically includes the following sub-steps: S410. Parallel Construction of Binary Classification Models: Based on the core predictor variable set, six different binary classification prediction models are simultaneously fitted on the same training set, including: Model 1: Logistic Regression Model. This model uses the sigmoid function to map linear combinations of features to probability outputs and employs L2 regularization to control model complexity. The predicted probability of this logistic regression model is expressed by the following formula: in: Used to characterize the sigmoid function; Model 2: Random Forest model, which integrates the prediction results of B=500 decision trees using the Bootstrap meta-algorithm. The sampling ratio of the feature subset of the random forest is set to the square root of the total number of features. The prediction probability of the random forest model is expressed by the following formula: Where: B is used to represent the number of decision trees, B=500; Used to characterize the The predicted probability of category 1 output by each decision tree; Model 3: Support Vector Machine model, which uses a radial basis function kernel to map low-dimensional features to a high-dimensional feature space, and then outputs the predicted probabilities scaled by Platt; the decision function is expressed as follows: in: Used to characterize the support vector index set; Used to characterize Lagrange multipliers; The class labels used to characterize the training samples correspond , correspond ; Used to characterize bias terms; Used to characterize radial basis kernel functions; Used to characterize the Euclidean distance between training samples and input samples; Used to characterize the kernel width parameter; Model 4: Extreme Gradient Boosting Model. Iteratively constructs a weak learner decision tree using a forward-step additive approach. Each tree fits the negative gradient of the current model to gradually reduce the prediction error. The prediction probability of the extreme gradient boosting model is expressed by the following formula: in: Used to characterize The cumulative output of the additive model after rounds of iteration; The sigmoid function is used to characterize the sigmoid function; T is used to characterize the number of decision trees; γ is used to characterize the learning rate. Model 5: Lightweight Gradient Boosting Machine Model, which discretizes feature intervals based on histogram algorithm and uses a leaf-growing strategy to find the optimal split point; Model Six: Naive Bayes Model. Based on Bayes' theorem and the assumption of conditional independence of features, it uses the Gaussian distribution assumption for continuous variables to estimate the posterior probability density. The predicted probability of the Naive Bayes model is expressed by the following formula: in: Used to characterize categories The prior probability; Used to characterize in category Next The conditional probability of each feature; S420. Grid Search Hyperparameter Configuration: For the six models of the binary classification prediction model, exhaustive combinations are performed within the preset candidate value range, and the classification performance verified by the fold cross-validation method with K=5 is used to determine the final configuration.

7. The method according to claim 2, characterized in that: Step S500 specifically includes the following sub-steps: S510. Quantitative Calculation of AUROC Discrimination: The true positive rate of each model of the binary classification prediction model at different classification thresholds is integrally expressed with respect to the false positive rate using the area under the receiver operating characteristic curve (AUC) assessment model, as follows: in: Used to characterize the The model uses the area under the receiver operating characteristic curve (AUC), which is dimensionless and takes values ​​of [value missing]. , When the value is 0.5, there is no distinguishing ability. Perfectly distinguishable when =1; Used to characterize the Each model at the threshold The true positive rate; Used to characterize the Each model at the threshold The false positive rate; Used to characterize the number of true positive cases; Used to characterize the number of false negatives; Used to characterize the number of false positives; Used to characterize the number of true negative cases; Used to characterize the probability of a classification threshold; S520. Quantitative Calculation of Calibration: The mean absolute error between the average predicted probability of each model in the binary classification prediction model and the actual observation frequency is calculated using the mean absolute error evaluation model, expressed by the following formula: in: Used to characterize the The mean absolute error of each model is dimensionless; G is used to characterize the number of calibration groups. Used to characterize the The model in the first The average predicted probability of the group; Used to characterize the The actual observation frequency of the group; S530. Quantitative Calculation of Clinical Practicality: The clinical practicality of each model of the binary prediction model is evaluated using the decision curve analysis and evaluation model, and the net benefit at each decision threshold is calculated and expressed by the following formula: in: Used to characterize the Each model at the threshold The net gain below is dimensionless; Used to characterize the Each model at the threshold The number of true positive cases; Used to characterize the Each model at the threshold The number of false positives; Used to characterize the size of the validation set; Used to characterize the probability of clinical decision thresholds; S540. Calculate the optimal classification cutoff threshold: The optimal classification cutoff threshold for each model of the binary classification prediction model is determined using the Youden index; the Youden index is expressed by the following formula: in: Used to characterize the The Yoden index of the model is dimensionless. Used to characterize the true negative rate; Used to characterize the optimal classification cutoff threshold; S550. Optimal Model Selection Decision: Among all candidate binary classification prediction models, the optimal prediction model is selected based on the following three quantitative logical judgment criteria, and its final performance index is locked: Validation set robustness judgment logic: Calculate the absolute value of the difference between the area under the receiver operating characteristic curve (AUC) of each binary classification prediction model on the training set and the area under the AUC of the receiver operating characteristic curve (AUC) on the validation set, and select the binary classification prediction model with the smallest absolute value of the difference; Optimal calibration judgment logic: Using the mean absolute error evaluation model, calculate the mean absolute error of each binary classification prediction model on the validation set, and select the binary classification prediction model with the lowest mean absolute error. Clinical applicability is highlighted by the following judgment logic: using the decision curve analysis and evaluation model, the net benefit value of the binary prediction model under each clinical decision threshold probability is calculated. Within the preset clinically commonly used threshold probability range, the net benefit value is selected to be higher than both the extreme intervention strategy line and the extreme non-intervention strategy line, and the net benefit value is higher than other binary prediction models. Based on the comprehensive comparison results of the quantitative logic, the logistic regression model is finally determined as the optimal prediction model, and the area under the receiver operating characteristic curve, the mean absolute error, and the optimal classification cutoff threshold corresponding to the Youden exponent maximization method are locked on the validation set.

8. The method according to claim 2, characterized in that: Step S600 specifically includes the following sub-steps: S610. General Score Mapping: Define the total scale range of the nodal plot, and construct the general score mapping formula based on the regression coefficients of each variable in the optimal prediction model, expressed as follows: in: Used to characterize the The scores of each variable in the nomogram; Used to characterize the total scale range of the nomogram; Used to characterize the maximum absolute value of all regression coefficients; Used to characterize the Measured values ​​of each variable; Used to characterize reference values, taking positive variables. or inverse variable ; Used to characterize the Normalization factor for the range of values ​​of each variable; S620. Execution of branch for differential mapping of three types of variables: Performing the forward mapping of the continuous variables is expressed as follows: ； Performing the inverse mapping of the continuous variables is expressed as follows: ； The binary variable mapping is performed as follows: ； S630. Construct the mapping relationship between the total score of the nodal plot and the intervention probability, expressed as follows: Wherein: TP is used to represent the total score of the nodal chart, and its value ranges from 0 to 100. , The dimensionless parameter is used to characterize the mapping from the total score of the nomogram to the intervention probability. Used to characterize the predictive probability of PTA intervention.

9. The method according to claim 2, characterized in that: Step S700 specifically includes the following sub-steps: S710. The continuous scores obtained by mapping the continuous score system of the nodal plot in step S600 are discretized into integer scores according to the parameter values ​​using the corresponding discretization algorithm; the discretization algorithm includes continuous variable forward discretization algorithm, continuous variable reverse discretization algorithm, and binary variable discretization algorithm. The continuous variable forward discretization algorithm is applicable to X. RI X Num X SII X DT Expressed as follows: The continuous variable inverse discretization algorithm is applicable to X. BAF X HR Expressed as follows: The binary variable discretization algorithm described above is applicable to X. TypeIV X PTAH X IT Expressed as follows: in: Used to characterize the The scores of each variable in the rating scale are non-negative integers. Used to characterize the The highest score of each variable; round is used to characterize the rounding function; S720. Quantitative division of weight levels: Based on the standardized weight coefficients, the core variables are divided into three levels: the first weight level includes the number of stenosis, resistance index, and brachial artery flow; the second weight level includes the systemic immune inflammation index, dialysis duration, heart rate, and type IV stenosis; the third weight level includes past PTA history and intimal thickening. S730. Risk stratification threshold calculation: Based on the sample distribution characteristics of the training set, the optimal binary classification threshold of 37 points is determined by the Youden exponent maximization method, and patients are divided into two levels: low risk (<37 points) and high risk (≥37 points).

10. The method according to claim 9, characterized in that: Step S800 specifically includes the following sub-steps: S810. Variable Collection and Quality Control: Collect the core predictive variable set and execute the following collection quality control rules: the dialysis duration is double-checked against the previous PTA history; the heart rate is limited to measurement before dialysis treatment and before the patient starts dialysis; the brachial artery flow, the resistance index, the number of stenosis, the type IV stenosis indicator variable, and the intimal thickening indicator variable are obtained by physicians with fixed qualifications; the systemic immune inflammation index is limited to blood routine results within two weeks. S820. Calculation of the total score of the rating scale: Substitute the collected measured values ​​into the corresponding discretization formula in step S710, calculate the scores of each item, and sum them to generate the patient's total score of the rating scale, expressed as follows: in: The patient's total score on the rating scale is a non-negative integer. S830. Calculation of Individual PTA Intervention Probability: Substitute the measured values ​​of the nine core predictive variables into the optimal prediction model to calculate the individual patient's PTA intervention prediction probability, expressed as follows: in: This is used to characterize the predicted probability of PTA intervention for the individual patient; the closer it is to 1, the more necessary PTA intervention is. S840. Automatic risk level determination: The total score of the rating scale is compared with the binary classification threshold and the risk is automatically classified into low-risk or high-risk levels; S850. Based on the risk level determination result of step S840, output the corresponding management strategy: if classified as low risk level, output the routine follow-up plan, with ultrasound re-examination every 3 months; if classified as high risk level, output the active assessment plan, with immediate comprehensive assessment, arrangement of PTA or surgical intervention, and multidisciplinary consultation.

Images