Patient internal fistula risk early warning method, system and equipment based on B ultrasonic data and medium

Abstract

The invention relates to a patient internal fistula risk early warning method, system and device based on B-mode ultrasound data and a medium. The method comprises the following steps: collecting B ultrasonic and clinical historical data of a patient and performing multi-modal preprocessing; function grading is achieved through an internal fistula grading model on the basis of B-mode ultrasound key features and converted into coding features; constructing an internal fistula risk causal graph through a causal discovery algorithm in combination with the coding features and clinical data; quantifying the causal effect between variables through a causal effect estimation method based on a causal graph; in combination with a causal diagram and an effect result, personalized risk early warning and attribution analysis results are generated through anti-factual reasoning, so that accurate and explainable internal fistula risk early warning is realized, and a basis is provided for clinical intervention.

Description

Technical Field

[0001] This invention belongs to the field of medical data processing technology, and in particular relates to a method, system, device and medium for early warning of patient arteriovenous fistula risk based on ultrasound data. Background Technology

[0002] With the widespread application of hemodialysis technology, the monitoring and evaluation of arteriovenous fistula (AVF) function has received increasing clinical attention. Currently, color Doppler ultrasound technology is mainly used to regularly examine AVFs, and the functional status of the fistula is assessed by obtaining vascular structure and hemodynamic parameters.

[0003] In traditional techniques, arteriovenous fistula (AVF) risk prediction mainly relies on doctors' visual interpretation of ultrasound images and clinical experience. In recent years, with the development of artificial intelligence (AI) technology, machine learning-based AVF risk assessment methods have emerged. These methods analyze ultrasound feature data to achieve automated grading and risk prediction of AVF function.

[0004] However, current early warning methods have significant limitations. First, traditional manual assessment methods are heavily influenced by physician experience, exhibiting strong subjectivity, and different operators may reach inconsistent conclusions. Second, while existing machine learning models can achieve automatic early warning, their internal decision-making process is like a "black box," unable to explain to physicians the specific causes and key influencing factors of the risk. More importantly, these models are susceptible to confounding factors, potentially misinterpreting statistical correlations as causal relationships, leading to reduced reliability of early warning results. These problems severely impact clinicians' trust in and willingness to adopt early warning results, limiting the widespread application of this technology in clinical practice. Summary of the Invention

[0005] Therefore, it is necessary to provide methods, systems, equipment, and media for early warning of patient arteriovenous fistula risk based on ultrasound data to address the aforementioned technical issues.

[0006] Firstly, this application provides a method for early warning of arteriovenous fistula risk in patients based on ultrasound data, including:

[0007] Ultrasound data and clinical history data of patients were collected. Multimodal preprocessing was performed on the ultrasound data and clinical history data to obtain a standardized multimodal dataset. The ultrasound data included structural data, hemodynamic data and examination metadata. The clinical history data included basic information, history of underlying diseases, dialysis-related data, history of complications and medication history.

[0008] Based on the key ultrasound features in the standardized multimodal dataset, the function of the arteriovenous fistula (AVF) in patients was graded using an AVF grading model, resulting in AVF functional grading results and coding features. The key ultrasound features include the ratio of peak systolic velocity, the rate of increase in vessel wall thickness, and the proportion of intimal rough areas.

[0009] Based on coding features and clinical historical data, a causal graph of arteriovenous fistula risk was constructed using a causal discovery algorithm, resulting in a causal graph structure representing the causal relationships between variables. The causal graph structure includes ultrasound indicators, patient baseline characteristics, treatment factors, and arteriovenous fistula functional classification as nodes.

[0010] Based on the causal graph structure, the magnitude of the causal effect among the variables is estimated using the causal effect estimation method, and the causal effect estimation results are obtained; the causal effect estimation results include direct causal effects and indirect causal effects.

[0011] Based on the causal graph structure and causal effect estimation results, personalized risk prediction is carried out through counterfactual reasoning, generating risk warning results and attribution analysis results. The risk warning results include risk level and warning type, and the attribution analysis results include main risk factors and causal paths.

[0012] In one embodiment, based on coding features and clinical history data, a causal graph of arteriovenous fistula risk is constructed using a causal discovery algorithm, resulting in a causal graph structure representing the causal relationships between variables, including:

[0013] Based on coding features and clinical history data, the partial correlation coefficients between variables were calculated through conditional independence tests to obtain the skeleton of the undirected graph;

[0014] Based on the undirected graph skeleton, the orientation rules are applied to determine the direction of the edges by time sequence constraints, resulting in a partially directed acyclic graph.

[0015] Based on a partially directed acyclic graph, the optimal causal graph structure is obtained by optimizing the Bayesian information criterion scoring function through a greedy equivalence search algorithm.

[0016] The causal sufficiency test is performed on the optimal causal graph structure, and unobserved confounding factors are identified through residual correlation analysis to obtain the causal graph structure.

[0017] In one embodiment, a causal sufficiency test is performed on the optimal causal graph structure, and unobserved confounding factors are identified through residual correlation analysis to obtain the causal graph structure, including:

[0018] Based on the optimal cause-effect graph structure, the residual term of each node is calculated through structural equation modeling to obtain the residual sequence;

[0019] For the residual series, the correlation strength between the residuals is calculated using the Pearson correlation coefficient, resulting in the residual correlation matrix; the formula for calculating the correlation strength is:

[0020]

[0021] in, Representing variables and The residual correlation coefficient, i.e., the correlation strength, Indicates the first The variable in the first... The residual values ​​of each sample Indicates the first The mean of the residuals of each variable, Indicates the number of samples;

[0022] Based on the residual correlation matrix, significant residual correlations are identified through hypothesis testing to obtain a set of potential confounding factors;

[0023] For a set of potential confounding factors, the causal graph structure is extended using latent variable modeling methods to obtain the causal graph structure.

[0024] In one embodiment, based on the causal graph structure, the magnitude of the causal effect among the variables is estimated using a causal effect estimation method to obtain the causal effect estimation result, including:

[0025] Based on the causal graph structure, the magnitude of direct causal effects is estimated through structural equation modeling to obtain regression coefficients;

[0026] Based on the regression coefficients, the magnitude of the causal effect is obtained by adjusting for confounding bias using a two-way stable estimation method.

[0027] For a binary treatment variable, the average treatment effect is estimated by propensity score matching to obtain the difference between the treatment group and the control group;

[0028] Based on the calibrated magnitude of the causal effect and the differences between the treatment and control groups, the causal effect estimation results are obtained.

[0029] In one embodiment, based on the regression coefficients, a two-way stable estimation method is used to adjust for confounding bias, resulting in the calibrated causal effect magnitude, including:

[0030] Based on the causal graph structure, a propensity score is calculated for each sample using a logistic regression model, resulting in a propensity score vector. The formula for calculating the propensity score is as follows:

[0031]

[0032] in, Indicates the first The propensity score of each sample. Indicates the variable being processed. Represents a covariate vector. Represents the intercept term. Represents the coefficient vector;

[0033] Based on the propensity score vector, the sample weights are adjusted using an inverse probability weighting method to obtain a weighted sample set;

[0034] For the weighted sample set, the regression coefficients are re-estimated using the weighted least squares method to obtain the preliminary calibrated causal effect;

[0035] Based on the preliminary calibrated causal effect, the calibrated causal effect size is obtained by combining dual robust estimation with regression adjustment and propensity score weighting.

[0036] In one embodiment, based on the causal graph structure and causal effect estimation results, personalized risk prediction is performed through counterfactual reasoning, generating risk warning results and attribution analysis results, including:

[0037] Based on the causal graph structure and the patient's current multimodal data, the posterior probability distribution of the node variables is calculated using the belief propagation algorithm to obtain the risk probability of arteriovenous fistula functional classification; the formula for calculating the risk probability is:

[0038]

[0039] in Indicates in the evidence Under certain conditions, the functional classification of arteriovenous fistula is as follows: The probability of a class Represents the normalization constant. Represents a node The potential function, Representing an edge The potential function, Indicates the number of nodes. Represents the set of evidence variables;

[0040] Based on risk probabilities, the intervention effect is simulated through counterfactual questioning, and the attribution weights of each risk factor are calculated to obtain the ranking of the main risk factors; the formula for calculating the attribution weights is:

[0041]

[0042] in, Indicates risk factors Attribution weights Indicates risk factors Conditional probability after intervention. Represents the conditional probability under the current observation conditions. This indicates the total number of risk factors. Indicates a reference value;

[0043] Based on the ranking of major risk factors, key causal chains from underlying diseases to changes in ultrasound indicators were extracted through causal path analysis, and the risk transmission mechanism was explained.

[0044] Based on risk probability, attribution weight, and risk transmission mechanism, a multi-level early warning strategy is used to generate risk early warning results and attribution analysis results. The multi-level early warning strategy triggers early warning signals of different colors according to the risk level. The risk level is divided based on the risk probability threshold.

[0045] Secondly, this application also provides a patient arteriovenous fistula risk early warning system based on ultrasound data, including:

[0046] The data acquisition and preprocessing module is used to collect patients' ultrasound data and clinical history data, and to perform multimodal preprocessing on the ultrasound data and clinical history data to obtain a standardized multimodal dataset. The ultrasound data includes structural data, hemodynamic data and examination metadata; the clinical history data includes basic information, history of underlying diseases, dialysis-related data, history of complications and medication history.

[0047] The data encoding and processing module is used to classify the function of the patient's arteriovenous fistula based on the key ultrasound features in the standardized multimodal dataset and through the arteriovenous fistula grading model, to obtain the arteriovenous fistula function grading results and encoded features; the key ultrasound features include the ratio of peak systolic velocity, the rate of increase of vessel wall thickness, and the proportion of intimal rough area.

[0048] The risk causal graph construction module is used to construct a causal graph of arteriovenous fistula risk based on coding features and clinical historical data, and obtain a causal graph structure representing the causal relationship between variables. The causal graph structure includes ultrasound indicators, patient baseline characteristics, treatment factors, and arteriovenous fistula functional classification as nodes.

[0049] The causal effect estimation module is used to estimate the magnitude of the causal effect between variables based on the causal graph structure and the causal effect estimation method, and obtain the causal effect estimation results; the causal effect estimation results include direct causal effects and indirect causal effects;

[0050] The risk warning and attribution analysis module is used to perform personalized risk prediction based on the causal graph structure and causal effect estimation results through counterfactual reasoning, and generate risk warning results and attribution analysis results. The risk warning results include risk level and warning type, and the attribution analysis results include main risk factors and causal paths.

[0051] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0052] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0053] The aforementioned method, system, equipment, and media for early warning of arteriovenous fistula (AVF) risk based on ultrasound data integrate multimodal ultrasound information, such as structural and hemodynamic data, with patient basic information and clinical history data, including medical history. A unified dataset is constructed through standardized preprocessing. An AVF grading model is used to quantitatively assess AVF function. Causal discovery algorithms and effect estimation methods are combined to clarify causal relationships and effect strengths between variables. Finally, personalized early warning and attribution analysis results are generated through counterfactual reasoning. This improves the accuracy of identifying AVF functional abnormalities, effectively avoiding the risks of missed and misdiagnosed diagnoses. By providing early warning of potential risks and identifying key risk factors and causal pathways, it provides medical staff with precise intervention basis, assisting in early intervention to extend the lifespan of AVFs and reduce the clinical need for secondary surgeries. The tiered early warning and personalized management strategies can optimize the allocation of medical resources, focusing resources on high-risk groups. Simultaneously, standardized data processing, tiered assessment, and early warning processes improve the quality of clinical services for AVF management. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart illustrating a patient arteriovenous fistula risk warning method based on ultrasound data in one embodiment.

[0056] Figure 2 This is a schematic diagram of a personalized early warning system for classifying and managing arteriovenous fistulas based on ultrasound data, as shown in a specific embodiment.

[0057] Figure 3 This is a schematic diagram of a patient arteriovenous fistula risk warning system based on ultrasound data in one embodiment. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] refer to Figure 1 The document presents a flowchart illustrating the patient arteriovenous fistula risk warning method based on ultrasound data provided in this application. The method includes the following steps:

[0060] S1: Collect the patient's ultrasound data and clinical history data, and perform multimodal preprocessing on the ultrasound data and clinical history data to obtain standardized multimodal data.

[0061] Specifically, the data acquisition process adopts a dual-path acquisition architecture of "direct device connection + system linkage". Ultrasound data is automatically acquired from clinical ultrasound equipment through a standardized interface that is compatible with mainstream standard formats and proprietary formats of various brands of mainstream equipment. Data parsing is achieved by integrating the development toolkit provided by the equipment manufacturer. Specifically, it includes structural data, hemodynamic data, and examination metadata. Structural data covers indicators such as the inner diameter of the arterial and venous ends, vessel wall thickness, intimal smoothness, and vessel course. Hemodynamic data includes parameters such as peak systolic velocity, end-diastolic velocity, blood flow direction, and blood flow continuity. Examination metadata includes examination time, operator information, equipment model, and examination site code. Clinical historical data is linked with the hospital information system, laboratory information system, and dialysis management system through a standardized medical data exchange protocol. After identity authentication is completed using a secure identity authentication protocol, multi-dimensional patient information is retrieved in batches. Basic information includes age, gender, weight, and time and type of arteriovenous fistula establishment. Basic medical history includes the duration of diabetes, hypertension classification, and cardiovascular disease diagnosis results. Dialysis-related data includes indicators such as dialysis frequency, single session duration, blood flow, and urea clearance index. Complication history records the occurrence time and treatment plan of problems such as arteriovenous fistula stenosis, thrombosis, and infection. Medication history covers the name, dosage, duration of medication, and compliance records of anticoagulants. The transmission of sensitive data is protected by encryption algorithms.

[0062] Multimodal preprocessing employs differentiated approaches based on the characteristics of the two types of data. Ultrasound data preprocessing includes three sub-steps: cleaning, standardization, and feature extraction. Data cleaning utilizes a dual filtering mechanism of "statistical verification + clinical rules." By calculating the mean and standard deviation of each parameter in historical normal samples, data exceeding the range of the mean plus or minus three times the standard deviation are marked as potential anomalies. A second verification is then performed in conjunction with clinical physiological ranges to remove invalid data caused by equipment malfunctions. Missing value imputation uses either linear interpolation at adjacent time points or estimation of the mean of samples matching the same features. Linear interpolation at adjacent time points is suitable for cases with a low missing value ratio, while estimation of the mean of samples matching the same features is used for scenarios with a high missing value ratio or without adjacent time point data. Format standardization is achieved by constructing an equipment parameter mapping table to unify the units, codes, etc., of different equipment into the system standard format, thus achieving data consistency and alignment. Key ultrasound feature extraction is based on clinical consensus to determine core indicators. The peak systolic velocity ratio is calculated by comparing the patient's current peak systolic velocity with the reference value for the corresponding fistula type. The fistula wall thickness growth rate is calculated using the wall thickness at the first examination after fistula establishment as the baseline value, by dividing the current thickness by the baseline value and then multiplying by 100%. The proportion of intimal irregularities is calculated by segmenting the intimal and lesion areas using an edge detection algorithm, and then using the ratio of the lesion area to the total intimal area. Clinical historical data preprocessing focuses on resolving the issues of coding and numerical consistency. Dumb variables are used for multi-category variables such as hypertension classification, and ordered variables such as medication adherence are converted to numerical values ​​using quantitative scoring. Continuous variables such as age and disease duration are standardized to eliminate the influence of dimensions. Finally, the processed ultrasound data and clinical data are associated with patient unique identifiers and examination times to form a standardized multimodal dataset containing multi-dimensional features.

[0063] S2, based on the key features of B-ultrasound in the standardized multimodal dataset, classifies the function of the patient's arteriovenous fistula using an arteriovenous fistula grading model, and obtains the grading results and coding features of the arteriovenous fistula function.

[0064] Specifically, a grading strategy combining machine learning models and clinical standards is employed to ensure the objectivity and clinical consistency of the grading results. The arteriovenous fistula (AVF) grading model utilizes the random forest algorithm, which boasts strong anti-overfitting capabilities and the ability to output feature importance. The model training data comes from labeled cases from multiple clinical institutions. Case inclusion criteria include a certain duration of AVF establishment, complete ultrasound data, and clinical diagnostic records; patients with serious underlying diseases are excluded. The model input features are key ultrasound features and auxiliary features. Key ultrasound features include the ratio of peak systolic velocity, the rate of increase in vessel wall thickness, and the proportion of intimal rough areas. Auxiliary features include the ratio of blood flow velocity and the rate of increase in vessel diameter. The output is a multi-level grading result, with grading standards strictly adhering to clinical guidelines, classifying AVFs into levels of normal function, mild abnormality, moderate abnormality, and severe abnormality, each corresponding to a clearly defined range of clinical indicators. Model training employs a reasonable dataset partitioning ratio, and algorithm parameters are optimized through grid search. Parameters include the number of decision trees, maximum depth, maximum number of features considered when splitting at each node, and minimum number of samples per leaf node, ensuring that the model performance meets clinical application requirements. After the classification is completed, one-hot encoding is used to convert the classification results into fixed-dimensional encoded features. These features will serve as one of the core variables for subsequent causal analysis. The encoding process is implemented through a professional machine learning toolkit to ensure that the encoding results are compatible with the model input format.

[0065] S3, based on coding features and clinical historical data, constructs a causal graph of arteriovenous fistula risk using a causal discovery algorithm, resulting in a causal graph structure representing the causal relationship between variables.

[0066] For example, the variable set for causal analysis is first determined, including ultrasound indicators, patient baseline characteristics, treatment factors, and vascular fistula functional grading coding features. Variable screening is completed through clinical expert review and mutual information testing, eliminating variables unrelated to fistula risk. The constraint-based PC algorithm is used for causal discovery. This algorithm constructs a causal graph by testing the conditional independence between variables. In practice, it is implemented using a professional data analysis toolkit. In the algorithm parameter settings, the independence test uses appropriate testing methods for continuous and categorical variables, setting reasonable significance levels and maximum condition set sizes to avoid test bias caused by high-dimensional conditions. The algorithm execution process consists of three stages. The first stage constructs an undirected graph, calculates the mutual information values ​​of all variable pairs, retains significantly correlated variable pairs, and constructs undirected connections. The second stage identifies the direction of some edges through V-structure identification. When a variable is independent of another variable in controlling a third variable, and the first variable is correlated with the third variable, and the third variable is correlated with another variable, the causal direction among these three variables is determined. The third stage optimizes the edge direction using Meek rules, eliminating loop structures and supplementing implicit causal relationships. To enhance the clinical validity of the causal graph, after the algorithm outputs preliminary results, a review panel composed of nephrology clinical experts and data scientists is invited to verify controversial causal relationships. Based on clinical knowledge, the direction of the edges is adjusted or spurious associations are deleted. In the final causal graph structure, nodes represent variables, directed edges represent causal relationships, and the edge labels indicate the results of independence tests to ensure reliability.

[0067] S4. Based on the causal graph structure, the magnitude of the causal effect among the variables is estimated using the causal effect estimation method, and the causal effect estimation result is obtained.

[0068] For example, causal effect estimation employs a combined approach of "propensity score matching + mediation analysis," selecting appropriate estimation models for different types of causal relationships. For cases where the treatment variable is binary and the outcome variable is continuous, propensity score matching is used. Propensity scores are calculated using a logistic regression model, incorporating covariates such as age and gender. Nearest neighbor matching is performed at an appropriate ratio with a matching tolerance. After matching, the balance of covariates is verified through statistical tests. Finally, the average treatment effect is calculated using the mean difference of the matched samples. The formula for calculating the average treatment effect is: , in the formula This represents the average treatment effect. This indicates the expected calculation. For the first The result variable values ​​when each sample is processed. For the first The outcome variable values ​​for samples not treated. For causal paths with mediating variables, mediation analysis was used to quantify direct and indirect effects. The total effect was decomposed using structural equation modeling. Model construction employed specialized statistical analysis software, and appropriate parameter estimation methods were used to estimate model parameters. The total effect equals the sum of direct and indirect effects, and the effect size is represented by standardized regression coefficients. Direct causal effects focus on the direct impact of the variable on the grading results, while indirect causal effects reflect the impact transmitted through mediating variables. All causal effect estimation results were validated for stability through sensitivity analysis. In propensity score matching, the matching ratio was adjusted, and in mediation analysis, the order of mediating variables was changed. If the effect size variation was within a reasonable range, the results were considered stable and reliable. The final output effect estimation results include the effect size, confidence interval, and test results.

[0069] S5, based on the causal graph structure and causal effect estimation results, performs personalized risk prediction through counterfactual reasoning, generating risk warning results and attribution analysis results.

[0070] Specifically, counterfactual reasoning is based on a structural causal model, which consists of a causal graph, variable functional relationships, and noise distribution. First, the effect estimation results of S4 are transformed into function parameters to construct a complete structural causal model. The variable functional relationships can be expressed as follows: , in the formula This represents the outcome variable, which can be an indicator such as the peak systolic velocity ratio. For the intercept term, The regression coefficients of the independent variables, i.e., the causal effect values, The independent variable could be a risk factor such as a marker of diabetes. The noise term follows a specific normal distribution. Personalized risk prediction involves inputting multimodal data from the target patient and completing the process through three steps: "factual reasoning → counterfactual intervention → outcome prediction." The factual reasoning stage calculates the probability of fistula grading in the patient's current state; the counterfactual intervention stage performs virtual adjustments to key risk variables while keeping other variables constant to eliminate interference; the outcome prediction stage calculates the grading probability after intervention using a structural causal model, comparing the probability changes before and after intervention to quantify the risk contribution of variables. Risk level classification combines grading results with risk probabilities, setting corresponding risk probability thresholds for different fistula grades. Risk levels are categorized as low, medium, and high risk, with different warning types corresponding to different risk level settings. Simultaneously, corresponding follow-up or intervention reminders are generated for different fistula grades. Attribution analysis results are sorted based on the magnitude of causal effects, extracting the leading risk factors and visually displaying the complete path from risk factors to fistula grading through a causal path diagram. All early warning and attribution information is synchronized to healthcare workers' workstations and patient terminals through a standardized data format. Healthcare workers' workstations display professional data and intervention suggestions, while patient terminals present interpretations and precautions in plain language. This allows healthcare workers to input intervention measures, automatically tracks data changes after intervention, updates risk prediction results, and forms a closed-loop management system.

[0071] The aforementioned arteriovenous fistula (AVF) risk warning method based on ultrasound data integrates multi-dimensional ultrasound data and clinical historical data, constructs a standardized dataset through multimodal preprocessing, and achieves quantitative assessment of AVF function based on an AVF grading model. It combines causal discovery algorithms and effect estimation techniques to reveal the intrinsic relationships between variables, and finally completes personalized risk prediction and attribution analysis through counterfactual reasoning. This effectively avoids the subjective bias of traditional assessments, improving the accuracy and reliability of AVF functional abnormality identification; through causal analysis and attribution interpretation, it provides medical staff with clear risk triggers and pathways, supporting targeted early intervention, thereby extending the lifespan of the AVF and reducing the physical and psychological burden of repeated surgeries for patients; and through grading assessment and personalized warnings, it achieves precise allocation of medical resources and improves clinical management efficiency.

[0072] To further illustrate the solutions of this application, a specific example is provided below. A personalized early warning system 200 for grading and managing arteriovenous fistulas based on ultrasound data includes:

[0073] 1. Ultrasound Data Acquisition Module 201: Establishes a data interface with clinical ultrasound equipment (supporting DICOM standard format and proprietary formats of mainstream ultrasound equipment), automatically acquiring multi-dimensional ultrasound data of the patient's arteriovenous fistula, including: Structural data: fistula diameter (arterial end, venous end), wall thickness, intimal smoothness (presence of plaques, calcification), vessel course (whether tortuous); Hemodynamic data: peak systolic velocity (PSV), end-diastolic velocity (EDV), blood flow direction, blood flow continuity (whether interrupted); Examination metadata: examination time, operating physician, ultrasound equipment model, examination site (e.g., radial artery-cephalic vein fistula in the forearm). Advantages: Eliminates the need for manual data entry, avoiding human error, and achieves real-time / near real-time acquisition of ultrasound data, providing timeliness assurance for subsequent analysis.

[0074] 2. Data Preprocessing Module 202: Cleans and standardizes the collected raw ultrasound data to eliminate data noise and format differences. Specific steps include: Data Cleaning: Removing outliers (such as data with a flow rate of 0 or far exceeding the physiological range due to equipment failure) and filling in missing values ​​(by interpolating data from adjacent time points or estimating the mean of data from patients of the same grade); Format Standardization: Converting non-standard data output from different ultrasound devices (such as inconsistent flow rate units: cm / s and m / s) into a unified clinical standard format; Feature Extraction: Extracting key features from structural and blood flow data, such as the ratio of PSV to the normal reference value, the growth rate of vessel wall thickness, and the proportion of intimal rough areas, to provide quantitative indicators for grading assessment.

[0075] 3. Patient History Database Module 203: Stores comprehensive medical data for patients, linking with the hospital's HIS (Hospital Information System), LIS (Laboratory Information System), and dialysis management system. Data includes: Basic information: age, gender, weight, fistula establishment time, fistula type (autogenous arteriovenous fistula, artificial fistula); Underlying medical history: duration of diabetes, hypertension classification, and presence of cardiovascular disease; Dialysis-related data: dialysis frequency, duration of each dialysis session, dialysis blood flow, and urea clearance index (Kt / V); Past complication history: history of fistula stenosis, thrombosis, infection, and aneurysm formation; Medication history: dosage and duration of anticoagulants (such as warfarin and low molecular weight heparin). Advantages: Provides individual patient background support for personalized early warning, enabling multi-dimensional analysis of ultrasound data and historical data.

[0076] 4. Arteriovenous fistula (AVF) grading assessment module 204: Based on pre-processed key ultrasound features, combined with a pre-defined AVF grading model, the module grades the patient's AVF function. Specifically, it includes: Grading model construction: Using machine learning algorithms (such as random forest and support vector machine), with clinically diagnosed AVF function cases (covering normal, mildly abnormal, moderately abnormal, and severely abnormal) as the training set, and using PSV ratio, vessel wall thickness, intimal smoothness, and presence of thrombus / stenosis as input features, an AVF grading model is constructed. Grading criteria: Grade 1 (Normal function): PSV within the normal reference range (autologous AVFs are usually 30-80 cm / s), vessel diameter meets standards (venous end ≥4 mm), vessel wall is smooth without plaque, and blood flow is continuous and uninterrupted; Grade 2 (Mildly abnormal): PSV slightly higher than the normal range (≤120 cm / s), or mild vessel wall thickening (thickness ≤0.5 mm compared to the first examination), no significant stenosis / thrombosis, and does not affect dialysis blood flow; Grade 3 (Moderately abnormal): PSV... Grade 120-300 cm / s (indicating moderate stenosis, 30%-50% stenosis), or the presence of small plaques (diameter ≤2 mm), with occasional fluctuations in dialysis blood flow (requiring adjustment of puncture position); Grade 4 (severe abnormality): PSV > 300 cm / s (indicating severe stenosis, stenosis > 50%), or the presence of thrombi (diameter > 2 mm), significantly decreased dialysis blood flow (< 200 ml / min), or risk of arteriovenous fistula occlusion. Advantages: Enables quantitative grading of arteriovenous fistula function, avoiding the subjectivity of manual assessment and providing a basis for subsequent grading management.

[0077] 5. Personalized Early Warning Analysis Module 205: Combining fistula grading results with patient historical database data, a patient-specific risk early warning model is constructed to predict the risk of fistula abnormalities. This includes: Risk Factor Association Analysis: Through regression analysis algorithms, the association between individual patient factors and fistula abnormalities is explored. For example, diabetic patients with PSV > 100 cm / s have a 2.3 times higher risk of thrombosis than non-diabetic patients; patients using low molecular weight heparin for < 3 months have a 1.8 times increased probability of mild abnormal fistulas progressing to moderate abnormalities; Risk Prediction: For different grades of fistulas, the abnormality risk is predicted within specific time windows (e.g., 1 week, 2 weeks, 1 month). For example, for Grade 1 fistulas: the focus is on predicting the risk of mild abnormalities due to vascular endothelial injury. Grade 1 abnormal risk is determined by factors such as the number of recent dialysis punctures (>3 punctures / week, increased risk); Grade 2 arteriovenous fistula: primarily predicts the risk of progression to moderate stenosis, considering whether the patient has hypertension (blood pressure >160 / 100 mmHg, increased risk); Grade 3 arteriovenous fistula: primarily predicts the risk of thrombosis, considering the patient's anticoagulant medication use (irregular medication use, increased risk); Grade 4 arteriovenous fistula: primarily predicts the risk of fistula occlusion, considering the patient's recent hemodynamic changes (PSV weekly increase >20 cm / s, increased risk); Warning threshold setting: personalized warning thresholds are set for each patient, such as for diabetic patients with Grade 2 arteriovenous fistulas, a warning is triggered when the weekly PSV growth rate >10 cm / s, rather than a uniform threshold.

[0078] 6. Early Warning Output and Management Module 206: Outputs grading results and early warning information in multiple formats and supports intervention record management for medical staff. Specifically, it includes: Graded Early Warning Signals: Outputs different types of early warnings based on risk level, such as no early warning for Grade 1 arteriovenous fistula (only periodic follow-up reminders), yellow early warning for Grade 2 arteriovenous fistula (reminding increased monitoring frequency), orange early warning for Grade 3 arteriovenous fistula (reminding follow-up ultrasound within 72 hours), and red early warning for Grade 4 arteriovenous fistula (reminding emergency intervention within 24 hours); Multi-Terminal Push: Early warning information is simultaneously pushed to medical staff terminals (such as hospital work phones and computer workstations, including basic patient information, key ultrasound images, and risk cause analysis) and patient mobile devices (such as WeChat official accounts and apps, including precautions such as avoiding lifting heavy objects with the arm on the side of the fistula and keeping it clean); Intervention Records: Supports medical staff in entering intervention measures (such as adjusting anticoagulant dosage, arranging angiography, and surgical intervention), and tracks changes in arteriovenous fistula grading after intervention, forming a closed loop of early warning-intervention-follow-up.

[0079] To further illustrate the solutions of the embodiments of this application, the following describes a specific embodiment applied to the above-mentioned personalized early warning system 200 for grading and managing arteriovenous fistulas based on ultrasound data, including the following steps:

[0080] Take, for example, a 58-year-old male patient with end-stage renal disease (autogenous radial artery-cephalic vein fistula, established 1 year ago, and with a 10-year history of type 2 diabetes):

[0081] Step 1: Ultrasound Data Acquisition: The patient undergoes an ultrasound examination of the arteriovenous fistula as instructed by the doctor. The system automatically acquires the following data through the DICOM interface with the GE Logiq E9 ultrasound device: Structural data: venous end diameter 4.2mm, arterial end diameter 3.8mm, wall thickness 1.2mm (0.3mm thicker than the first examination), with a 1.5mm plaque in the intima; Hemodynamic data: PSV 115cm / s, EDV 45cm / s, normal blood flow direction, continuous and uninterrupted; Examination metadata: Date: 202X, Physician: Zhang XX, Examination site: forearm arteriovenous fistula.

[0082] Step 2: Data Preprocessing: The system processes the collected data: Data Cleaning: Confirms that data such as PSV 115cm / s and wall thickness 1.2mm are normal and do not need to be discarded; Format Standardization: Converts the "PSV unit m / s" output by the ultrasound equipment to the clinical standard "cm / s"; Feature Extraction: Calculates "PSV ratio = 115 / 60 ≈ 1.92 (normal reference value is 60cm / s)" and "Wall thickness growth rate = 0.3 / 0.9 ≈ 33.3% (wall thickness of 0.9mm in the first examination)".

[0083] Step 3: Fistula grading assessment: The system inputs the pre-processed features (PSV ratio 1.92, wall thickness 1.2mm, plaque diameter 1.5mm) into the grading model: Model calculation: PSV 115cm / s (≤120cm / s), slight wall thickening, presence of small plaques (≤2mm), meeting the "mild abnormality" standard; Output grading result: Fistula functional level is grade 2.

[0084] Step 4: Personalized Early Warning Analysis: The system retrieves historical patient data: The patient has had diabetes for 10 years, with dialysis punctures 4 times / week in the past month, and irregular use of anticoagulants (prescribed low molecular weight heparin 4000 IU / day, actually used 2-3 times / week); Risk Association Analysis: Combining the characteristics of a grade 2 arteriovenous fistula with the patient data, the system uses a regression model to determine that "due to diabetes + frequent punctures + irregular anticoagulation, the patient's risk of grade 2 arteriovenous fistula progressing to grade 3 (moderate stenosis) is 2.1 times higher than that of ordinary patients"; Risk Prediction: The probability of progression to grade 3 arteriovenous fistula within the next 2 weeks is predicted to be 65%, triggering a yellow warning.

[0085] Step 5: Early Warning Output and Intervention: Multi-terminal push notification: Medical staff receive a yellow warning at their workstation (including patient data, risk reasons, and recommendations to "increase ultrasound monitoring frequency to once a week and strengthen anticoagulation medication education"); the patient receives a notification on their WeChat official account ("Your arteriovenous fistula is currently mildly dysfunctional. Please avoid lifting heavy objects with the arm on the side of the fistula, use anticoagulants as prescribed, and contact medical staff immediately if the tremor weakens"); Intervention and follow-up: Medical staff educate the patient on medication, and the patient begins to use anticoagulants regularly; a follow-up ultrasound examination one week later shows that the PSV has decreased to 95 cm / s, the fistula wall thickness has not increased, the system reassesses the fistula grade as still level 2, the warning is lifted, and routine monitoring (ultrasound every 2 weeks) is resumed.

[0086] Through the above implementation process, the system achieves accurate grading and personalized early warning of the patient's arteriovenous fistula, effectively preventing further deterioration of the fistula function.

[0087] In an optional embodiment, based on coding features and clinical history data, a causal graph of arteriovenous fistula risk is constructed using a causal discovery algorithm to obtain a causal graph structure representing the causal relationships between variables, including the following steps:

[0088] S11, based on coding features and clinical historical data, calculates the partial correlation coefficients between variables through conditional independence tests to obtain the skeleton of the undirected graph.

[0089] Specifically, conditional independence testing is fundamental for identifying potential associations between variables. Its core logic is to determine whether a statistical association still exists between two target variables after controlling for a third variable. In practice, the variable set, composed of coded features and clinical historical data, is first categorized by data type. Continuous variables, such as age and disease duration, are tested using the Gaussian conditional independence test; categorical variables, such as gender and hypertension classification, are tested using the chi-square conditional independence test; and mixed-type variable combinations are tested using a combination of kernel density estimation. During the test, the partial correlation coefficient quantifies the strength of the association between variables. The calculation of the partial correlation coefficient is performed by controlling for different numbers of conditional variables, starting with zero variables (simple correlation) and gradually increasing the number of conditional variables to a preset maximum value. The maximum value is usually set to the square root of the total number of variables to balance test precision and computational efficiency. When the absolute value of the partial correlation coefficient is less than a preset threshold, the two variables are considered independent under the current conditions, and the undirected connection between them is deleted; otherwise, the connection is retained. This threshold is determined by the significance level α, typically with α = 0.05 corresponding to a threshold of approximately 0.08. The final result is an undirected graph skeleton containing only significantly related variable pairs. Each node in the skeleton represents a variable, and the undirected edges between nodes represent potential causal relationships.

[0090] S12, based on the undirected graph skeleton, uses time sequence constraints to apply orientation rules to determine the direction of edges, resulting in a partially directed acyclic graph.

[0091] Specifically, the undirected graph skeleton only reflects variable relationships and needs to be transformed into a structure with causal orientation through direction determination. Time sequence constraints are the most direct basis for direction determination, using the order in which variables occur or are measured to determine the causal flow; a variable that occurs earlier may become the cause of a variable that occurs later, but the reverse is not true. In the variable set of this method, variables such as the history of underlying diseases and the time of arteriovenous fistula establishment have clear time attributes and can be directly used as constraints. Data on the history of underlying diseases, such as the time of diagnosis of diabetes, belongs to this category. Taking the duration of diabetes as an example, its measurement time is earlier than the rate of increase in fistula wall thickness; therefore, the direction of the edge is determined to be from the duration of diabetes to the rate of increase in fistula wall thickness. For variable pairs without a clear time sequence, the PSV ratio and fistula grading belong to this category. Orientation rules are applied for further judgment, with core rules including V-structure identification and Meek rules. V-structure identification targets a group of nodes composed of three variables. If variables A and C are independent in controlling B, and A and B, and B and C are both related, then a V-structure of A→B←C is formed, with B as the common cause of A and C. Meek rules are used to extend direction determination. For example, if there are edges AB and B→C, and A and C are not directly connected, then AB is directed to A→B to avoid forming a cycle. By combining time constraints and orientation rules, most undirected edges are given a clear direction, while the remaining edges whose direction cannot be determined remain undirected, forming a partially directed acyclic graph.

[0092] S13. Based on a partially directed acyclic graph, the optimal causal graph structure is obtained by optimizing the Bayesian information criterion scoring function through a greedy equivalence search algorithm.

[0093] Specifically, some directed acyclic graphs still exhibit structural uncertainty, requiring a scoring search algorithm to find the structure that best fits the data. The greedy equivalence search algorithm is a commonly used scoring search method in causal discovery. Its core is to traverse possible structures within causal equivalence classes, evaluate the fit between the structure and the data using a scoring function, and finally select the structure with the optimal score. This step uses the Bayesian information criterion as the scoring function, which considers both the fit of the structure to the data and the structural complexity. The calculation formula is as follows: . In the formula, This is the Bayesian information criterion score; the smaller the score, the better the fit between the structure and the data. The log-likelihood value reflects how well the structure fits the data; the larger the value, the better the fit. This refers to the number of model parameters, i.e., the number of edges in the causal graph. For sample size, This term penalizes overly complex structures to avoid overfitting. During algorithm execution, it starts from a partially directed acyclic graph and generates adjacent structures through operations such as adding, deleting, and reversing edges, calculating the value of each structure. The scoring process involves selecting the structure with the highest score as the current optimal structure and repeating this process until the score can no longer be improved through local adjustments. The structure at this point is the optimal cause-effect graph structure, which maintains simplicity while fitting the data.

[0094] S14. Perform a causal sufficiency test on the optimal causal graph structure, identify unobserved confounding factors through residual correlation analysis, and obtain the causal graph structure.

[0095] For example, the optimal causal graph structure may be biased due to the omission of unobserved confounding factors. Variables that are difficult to quantify, such as patient genetic background and lifestyle habits, may fall into this category. The causal adequacy test aims to identify these potential factors and improve the structure. This test is achieved through residual correlation analysis. The core logic is that if the causal graph fully captures all important causal relationships, then the residuals of the variables—that is, the parts that cannot be explained by the model—should be independent of each other. If there is a significant correlation between the residuals, it suggests the presence of unobserved confounding factors affecting these two variables. The specific implementation consists of four sub-steps: calculating residuals using structural equation modeling, quantifying residual correlations, identifying potential confounding factors, and expanding the causal graph to ensure the adequacy and reliability of the final causal graph structure.

[0096] In an optional embodiment, the optimal causal graph structure is subjected to a causal sufficiency test, and unobserved confounding factors are identified through residual correlation analysis to obtain the causal graph structure, including the following steps:

[0097] S21, based on the optimal causal graph structure, calculates the residual term of each node through structural equation modeling to obtain the residual sequence.

[0098] Specifically, the optimal cause-effect graph structure clearly defines the direct causal relationships between variables, while structural equation modeling quantifies these relationships into mathematical equations, providing a foundation for calculating residuals. When constructing a structural equation model, each node variable is expressed as the sum of a linear combination of its parent node variables and the residual term, in the following form: ,in For the intercept term, to The regression coefficients of the parent node variables. That is The residual terms represent the parent node variables that cannot be interpreted. The variability component also signals the potential presence of confounding factors. The model parameters are solved using maximum likelihood estimation, minimizing the deviation between observed values ​​and model predictions to obtain optimal estimates of the intercept and regression coefficients. Subsequently, the observed values ​​of each variable and the parent node variable are substituted into the equation to calculate the residual value of each variable for each sample. For example, for the ... One sample, variables The residual values: The residual values ​​of all variables are arranged in the order of the samples to form the corresponding residual sequence.

[0099] S22, For the residual series, the correlation strength between residuals is calculated using the Pearson correlation coefficient to obtain the residual correlation matrix; the formula for calculating the correlation strength is:

[0100]

[0101] in, Representing variables and The residual correlation coefficient, i.e., the correlation strength, Indicates the first The variable in the first... The residual values ​​of each sample Indicates the first The mean of the residuals of each variable, Indicates the number of samples.

[0102] Specifically, in an ideal causal graph structure, the residuals of all variables should be independent of each other, because the residuals represent the variation that the parent node cannot explain. If there is a significant correlation between the residuals, it indicates that there may be unobserved confounding factors affecting both variables simultaneously. The Pearson correlation coefficient is a classic indicator of the strength of the linear correlation between two residual sequences. In the above formula, Representing variables and The residual correlation coefficient ranges from [-1, 1], and the larger the absolute value, the stronger the correlation. Indicates the first The variable in the first... The residual values ​​of each sample Indicates the first The mean of the residuals of each variable, This represents the sample size. During the calculation, the mean of each residual sequence is first calculated, then the covariance is calculated in the numerator, and the standard deviation is calculated in the denominator, resulting in the correlation coefficient. The residual correlation coefficients between all pairs of variables are arranged in matrix form, with rows and columns corresponding to the variables, and matrix elements representing the correlation coefficients of the corresponding variable pairs. This forms the residual correlation matrix, providing data support for subsequent identification of potential confounding factors.

[0103] S23, based on the residual correlation matrix, identifies significant residual correlations through hypothesis testing to obtain a set of potential confounding factors.

[0104] Specifically, the magnitude of the residual correlation coefficient only reflects the strength of the correlation. It needs to be determined through statistical hypothesis testing to determine whether the correlation is statistically significant, rather than being caused by random error. The null hypothesis for the test is that the residuals of the two variables are not significantly correlated (i.e., the correlation coefficient is not statistically significant). The alternative hypothesis is that the residuals of the two variables are significantly correlated (i.e., ).use The test statistic is calculated using the following formula to perform the hypothesis test: ,in Where is the sample size, this statistic follows a set of degrees of freedom. of Distribution. Set significance level. (Usually taken as 0.05), if the calculated result The statistic corresponds to Value less than If the correlation is significant, the null hypothesis is rejected, indicating a significant correlation between the residuals of the two variables. This test is performed on each off-diagonal element of the residual correlation matrix to collect all variable pairs with significant residual correlation. Subsequently, the optimal causal graph structure is used for screening: if the two variables have no direct or indirect causal path in the graph, but their residuals are significantly correlated, then an unobserved confounding factor is identified that simultaneously affects both variables, and this potential confounding factor is included in the set; if the correlation between the residuals of the two variables can be explained by existing causal paths (e.g., the existence of a common parent node), then the variable pair is excluded, ultimately forming a set containing only the true potential confounding factors.

[0105] S24. For the set of potential confounding factors, the causal graph structure is extended by the latent variable modeling method to obtain the causal graph structure.

[0106] Specifically, latent confounding factors are latent variables that are not directly observed and need to be incorporated into the causal graph through statistical modeling to improve the expression of causal relationships. A factor analysis model is used to model the latent variables. This model assumes that the latent variables (latent confounding factors) are the common cause of the correlation between the residuals of multiple observed variables, and the latent variables are characterized by extracting common factors from the residual sequences. During the modeling process, the number of latent variables is first determined based on the number of significantly correlated pairs of variables. Typically, one latent variable can explain a set of mutually correlated pairs of variables, avoiding excessive latent variables that would complicate the model. Then, the factor loading matrix is ​​solved using maximum likelihood estimation. The magnitude of the loading coefficients reflects the strength of the influence of the latent variables on the observed variables. Observed variables strongly correlated with the latent variables are selected based on the absolute value of the loading coefficients (usually with a threshold of 0.5). Based on the original optimal causal graph structure, latent variable nodes representing latent confounding factors (labeled with U) are added, and directed edges pointing from the latent variables to the corresponding observed variables are added according to the direction of the factor loadings. The weight of the edges is quantified by the loading coefficients. The expanded causal graph needs to undergo a cycle-free test again to ensure that no directed loop structure is formed. Finally, the BIC value of the expanded causal graph is recalculated to verify whether its goodness of fit is better than the original structure. If the goodness of fit is improved and the significant correlation terms in the residual correlation matrix disappear, then the expanded structure is determined to be the final causal graph structure, which can be used for subsequent causal effect estimation.

[0107] In an optional embodiment, based on the causal graph structure, the magnitude of the causal effect among the variables is estimated using a causal effect estimation method to obtain the causal effect estimation result, including the following steps:

[0108] S31, based on the causal graph structure, estimates the magnitude of direct causal effects through structural equation modeling to obtain regression coefficients.

[0109] Specifically, a direct causal effect refers to the net effect of one variable on another after controlling for all its mediating variables. This effect can be directly quantified by the regression coefficients in a structural equation model. Structural equation models are constructed based on causal graphs, where the equation for each node variable corresponds to its parent node in the causal graph. If the variable... The parent node is , ,but The structural equation is . In the formula, and That is and right The regression coefficient of the direct causal effect, This is the intercept term, used to adjust the base level of the variable. The residuals represent the portion of variation that the model cannot explain. For continuous outcome variables, the PSV ratio falls into this category, and ordinary least squares (OLS) is used to estimate the regression coefficients. For categorical outcome variables, the arteriovenous fistula grade falls into this category, and generalized least squares (GLS) combined with the appropriate link function, such as ordered logit linking, is used. The sign of the regression coefficients indicates the direction of the causal effect; positive coefficients indicate a promoting effect, and negative coefficients indicate an inhibiting effect. The absolute value represents the strength of the effect. For example... express For each additional unit, An average increase of 0.3 units is required; at this point, control is necessary. The values ​​remain unchanged. This method yields regression coefficients for all direct causal pathways, providing preliminary results for effect quantification.

[0110] S32, based on the regression coefficients, adjusts for confounding bias using a two-way stable estimation method to obtain the calibrated causal effect size.

[0111] Specifically, based on the regression coefficients, a two-way stable estimation method is used to adjust for confounding bias, resulting in the calibrated causal effect size. Initial regression coefficients may be affected by incompletely controlled confounding factors. For example, in the association between anticoagulant dosage and arteriovenous fistula thrombosis risk, patient compliance, an incompletely observed variable, may simultaneously affect both dosage and thrombosis occurrence, leading to coefficient estimation bias. The core advantage of the two-way stable estimation method lies in balancing inter-group confounding through propensity score calculations and controlling residual bias through the regression model, overcoming the limitations of single adjustment methods. This method first uses a causal graph to screen core confounding variables, including only those that simultaneously affect both treatment and outcome variables, constructing a propensity score model and a regression adjustment model. Subsequently, the initial regression coefficients are calibrated using inverse probability weighting and regression coefficient correction. The final calibrated effect size needs to be validated through sensitivity analysis, such as changing the scope of confounding variables to ensure stability even with potential unobserved confounding, providing an unbiased basis for subsequent effect integration.

[0112] S33. For a binary treatment variable, the average treatment effect is estimated by propensity score matching to obtain the difference between the treatment group and the control group.

[0113] Specifically, a binary treatment variable refers to a variable that takes only a yes or no value, such as whether or not one has diabetes or whether or not one regularly takes anticoagulants. The causal effect of this type of variable can be quantified by the average treatment effect, which is the difference in outcome variables between the treatment group (samples with a yes value) and the control group (samples with a no value). The core of propensity score matching is to make the treatment group and the control group comparable in terms of confounding variables through matching, thereby attributing the differences between groups to the causal effect of the treatment variable. In practice, firstly, the propensity score of each sample is calculated, which is the probability of the sample receiving the treatment. Then, based on the propensity score, the nearest neighbor matching method is used to match 1-4 control group samples to the treatment group samples. The matching tolerance is set to 0.2 times the standard deviation of the propensity score to ensure the quality of the matching. After matching is completed, the mean difference of the outcome variables between the treatment group and the control group is calculated. This mean difference is the average treatment effect. For example, if the mean difference of the PSV ratio between the group with diabetes and the group without diabetes is 0.25, it indicates that the average treatment effect of diabetes on the PSV ratio is 0.25.

[0114] S34. Based on the calibrated causal effect magnitude and the difference between the treatment group and the control group, the causal effect estimation results are obtained.

[0115] Specifically, the calibrated direct causal effect and the average treatment effect of the binary treatment variables reflect the strength of the causal relationship from different dimensions, and they need to be integrated to form a complete effect estimation result. During the integration process, the consistency of the two effect results is first tested. If the regression coefficients of the direct effect and the average treatment effect have the same sign and similar values, with a relative difference of <20%, the result is considered reliable. If a difference exists, the model specification and parameter selection are re-examined to eliminate method application errors. The final output of the causal effect estimation result includes three parts: the direct causal effect is expressed as the calibrated regression coefficient with a 95% confidence interval; the indirect causal effect is obtained by subtracting the direct effect from the total effect, and the total effect is calculated using the regression coefficients of the uncontrolled mediator variable; the average treatment effect of the binary treatment variable is expressed as the difference between the group means with a p-value. These contents provide a comprehensive quantitative basis for subsequent risk prediction.

[0116] In an optional embodiment, based on the regression coefficients, the confounding bias is adjusted using a two-way stable estimation method to obtain the calibrated causal effect magnitude, including the following steps:

[0117] S41, based on the causal graph structure, calculates the propensity score for each sample using a logistic regression model, obtaining a propensity score vector; the formula for calculating the propensity score is:

[0118]

[0119] in, Indicates the first The propensity score of each sample. Indicates the variable being processed. Represents a covariate vector. Represents the intercept term. This represents the coefficient vector.

[0120] Specifically, the core function of propensity score is to quantify the probability of a sample receiving a treatment (such as a specific therapeutic intervention or exposure to a risk factor), used to balance the distributional differences in confounding variables between the treatment and control groups. In the above formula, where... Indicates the first The propensity score of each sample, with a value range of [0,1]; This represents the treatment variable, with a value of 1 indicating that the sample belongs to the treatment group and 0 indicating the control group. The covariate vector must be chosen strictly according to the cause-effect graph structure and only include variables relevant to the treatment variables. Confounding variables that have a causal effect on both the outcome variable (risk of arteriovenous fistula) and the mediating variable should be avoided to prevent bias caused by including mediating or collision variables. This represents the intercept term of the logistic regression model. This represents the coefficient vector of the covariates, reflecting the strength of each confounding variable's influence on the propensity score. During model construction, these variables are considered. The dependent variable and the covariate vector Using maximum likelihood estimation as the independent variable, the solution is obtained. and The optimal value is determined to ensure the model can most accurately predict the propensity score for each sample. For each sample, its covariate observations are substituted into the trained logistic regression model to calculate the corresponding propensity score. All samples Arranging them in order forms a tendency score vector, which provides a basis for subsequent sample weight adjustments.

[0121] S42, based on the propensity score vector, adjusts the sample weights using an inverse probability weighting method to obtain a weighted sample set.

[0122] In observational studies, the distribution of confounding variables between the treatment and control groups is often unbalanced. Directly comparing the results of the two groups can introduce confounding bias. Inverse probability weighting (IPTW) assigns a weight to each sample that is inversely proportional to the propensity score, making the distribution of confounding variables in the weighted sample set more consistent between the treatment and control groups, thus simulating the effect of a randomized controlled trial. The weighting calculation rules are clear: the treatment group sample ( The weight of ) is control group sample ( The weight of ) is The logic behind this rule is: bias towards scoring. Smaller treatment group samples (i.e., those that were "unlikely" to receive treatment but actually did) were given greater weight and tended to receive higher scores. Larger treatment group samples are assigned smaller weights; similarly, control group samples with propensity scores close to 1 (i.e., samples that were likely to receive treatment but did not) are assigned larger weights. After weight calculation, the weights need to be standardized by dividing the weight of each sample by the mean of all sample weights to avoid model instability due to excessive weight differences. The original sample data are then correlated with the corresponding standardized weights to form a weighted sample set. In this set, confounding variables are balanced between the treatment and control groups, laying the foundation for estimating unbiased causal effects.

[0123] S43. For the weighted sample set, the regression coefficients are re-estimated using the weighted least squares method to obtain the preliminary calibrated causal effect.

[0124] For example, weighted least squares is an extension of ordinary least squares. Its core is the introduction of sample weights into the regression calculation, allowing samples with higher weights to contribute more significantly to the estimation of regression coefficients. Substituting the weighted sample set into the original structural equation model, using the sample weights as the weight matrix, the weighted least squares method is used to re-estimate the regression coefficients of the treatment variable on the outcome variable. Compared to ordinary least squares, this method highlights the role of representative samples through weight adjustment, reducing the bias caused by the unbalanced distribution of confounding variables. For instance, in estimating the effect of hypertension on the PSV ratio, after balancing confounding variables such as age and disease duration using inverse probability weighting, the re-estimated regression coefficients represent the preliminary calibrated direct causal effect. This effect has eliminated spurious associations caused by confounding factors and is closer to the true strength of the causal relationship.

[0125] S44. Based on the preliminary calibrated causal effect, the calibrated causal effect size is obtained by combining dual robust estimation with regression adjustment and propensity score weighting.

[0126] Specifically, dual robust estimation is a key step in further improving the reliability of effect estimation. Its core is the simultaneous application of regression adjustment and propensity score weighting methods. If the assumptions of one method hold true—for example, the propensity score model or the regression model is correctly specified—the overall estimation result will be consistent. In practice, first, a regression model is constructed including the treatment variable, confounding variables, and their interaction terms. Ordinary least squares is used to estimate the regression coefficients of the treatment variable, i.e., the regression adjustment effect. Then, combined with the weighted regression coefficients obtained in S43, i.e., the propensity score weighted effect, the dual robust effect estimate is calculated using a weighted average. The weights are determined based on the goodness of fit of the two methods, such as the R² value; the method with higher goodness of fit has a larger weight. This approach combines the advantages of both methods, reducing the risk of errors in the specification of a single method. The final dual robust effect estimate is the calibrated causal effect size, possessing both accuracy and robustness, and can serve as a core quantitative indicator for subsequent risk analysis.

[0127] In an optional embodiment, based on the causal graph structure and causal effect estimation results, personalized risk prediction is performed through counterfactual reasoning to generate risk warning results and attribution analysis results, including the following steps:

[0128] S51, based on the causal graph structure and the patient's current multimodal data, calculates the posterior probability distribution of node variables using the belief propagation algorithm to obtain the risk probability of arteriovenous fistula functional classification; the formula for calculating the risk probability is:

[0129]

[0130] in Indicates in the evidence Under certain conditions, the functional classification of arteriovenous fistula is as follows: The probability of a class Represents the normalization constant. Represents a node The potential function, Representing an edge The potential function, Indicates the number of nodes. This represents the set of evidence variables.

[0131] Specifically, the belief propagation algorithm is a message-passing algorithm for efficiently calculating the posterior probability of variables in graph models. It is suitable for directed acyclic graph structures such as causal graphs. Its core is to achieve local calculation of global probabilities by passing "belief messages" between adjacent nodes. In the above formula for calculating risk probability, Indicates in the evidence (Based on the patient's current ultrasound data, clinical history data, etc.) the functional classification of the arteriovenous fistula is as follows: The probability of a class For the functional grading of arteriovenous fistulas, For specific classification categories; This is a normalization constant used to ensure that the sum of the probabilities of all hierarchical categories is 1. Its value is obtained by summing the numerator over all hierarchical categories. For nodes The potential function for evidence nodes (variables for which the patient currently has observational data). For latent variable nodes, the probability of the observed value is taken directly; for latent variable nodes, the prior probability is taken. For the edge The potential function is used to quantize the nodes. and The strength of the association between nodes is determined by the weights (magnitude of causal effect) of the edges in the causal graph. During algorithm execution, the patient's current data is first passed as evidence to the corresponding node. Then, likelihood messages are passed from the leaf nodes to the root node (the fistula functional grading node) of the causal graph, followed by prior messages from the root node to the leaf nodes. The marginal probability of each node is calculated through message fusion, ultimately yielding the fistula functional grading variable. In evidence The posterior probability distribution under the given conditions, the probability value of each category is the corresponding risk probability.

[0132] S52, based on risk probability, simulates the intervention effect through counterfactual questioning, calculates the attribution weight of each risk factor, and obtains the ranking of the main risk factors; the formula for calculating the attribution weight is:

[0133]

[0134] in, Indicates risk factors Attribution weights Indicates risk factors Conditional probability after intervention. Represents the conditional probability under the current observation conditions. This indicates the total number of risk factors. Indicates a reference value.

[0135] Specifically, counterfactual questioning is a core tool of causal reasoning, used to answer the question of how a patient's risk of arteriovenous fistula would change if a certain risk factor changed. Its essence is to separate the independent effects of the factor by intervening in the values ​​of the variable. In the formula for calculating the attribution weights mentioned above, Indicates risk factors The attribution weight, with a value range of [0,1], indicates that the larger the weight, the greater the contribution of the factor to the patient's current arteriovenous fistula risk; Indicates risk factors To implement idealized intervention, its value is set as a reference value. Reference values ​​are typically the clinically normal range for that factor (such as the normal threshold for blood pressure, the standard dose of anticoagulants, etc.). During the intervention, the values ​​of other variables should be kept constant to ensure that only the values ​​of the factors are reflected. Independent influence; Risk probability for functional classification of arteriovenous fistula after intervention (usually the probability of the high-risk category). The absolute value of the difference between the two values ​​represents the high-risk probability under the patient's current observation conditions. The contribution of each risk factor to the overall risk is calculated. The denominator is the sum of the contributions of all risk factors, used to standardize the attribution weights and ensure that the sum of the weights of all factors is 1. During the calculation, counterfactual intervention and probability calculation are performed on each risk factor one by one to obtain the corresponding attribution weight. These weights are then sorted from largest to smallest to form a list of major risk factors, providing a basis for determining the priority of clinical interventions.

[0136] S53, based on the ranking of major risk factors, extracts the key causal chain from the underlying disease to the changes in ultrasound indicators through causal path analysis, and obtains an explanation of the risk transmission mechanism.

[0137] Specifically, while the ranking of major risk factors identifies key factors, it doesn't reveal how these factors influence fistula function through causal relationships. Causal path analysis aims to fill this gap by tracing directed paths in the causal graph to construct a complete risk transmission chain. The analysis starts with the major risk factors and ends with the fistula functional grading, extracting all directed paths from the start to the end in the causal graph. Key paths are selected based on the magnitude of the causal effects along these paths. The larger the product of the regression coefficients of all direct effects along a path—the greater the path effect—the more critical the path. For example, if the major risk factor is the duration of diabetes, the extracted key path might be: duration of diabetes → vascular endothelial damage → increased rate of increase in vessel wall thickness → progression of fistula grading. The path effect is the product of the direct effect coefficients of each link. A mechanistic explanation is provided for each key path, combining clinical knowledge to explain the biological or medical principles of each link. For instance, prolonged diabetes leads to chronically elevated blood glucose, damaging vascular endothelial cells, reducing vessel wall repair function, leading to vessel wall thickening, and ultimately increasing the risk of fistula abnormalities. By integrating multiple key paths, a comprehensive explanation of the risk transmission mechanism is formed, enabling healthcare professionals to understand the intrinsic logic of risk generation.

[0138] S54, based on risk probability, attribution weight, and risk transmission mechanism, generates risk warning results and attribution analysis results through a multi-level early warning strategy; the multi-level early warning strategy triggers warning signals of different colors according to the risk level; the risk level is divided based on the risk probability threshold.

[0139] Specifically, the core of the multi-level early warning strategy is to combine risk probability with clinical intervention needs, classify risk levels and trigger corresponding early warnings, and combine attribution information with early warning results to provide a clear direction for intervention. Risk levels are based on risk probability thresholds, which are determined with reference to clinical intervention guidelines. For example, when an arteriovenous fistula is classified as mildly abnormal, a risk probability <20% is low risk, 20%-40% is medium risk, and >40% is high risk. Different risk thresholds exist for different levels, with lower thresholds for more severe levels to facilitate early intervention. The multi-level early warning strategy triggers different colored warning signals based on the risk level: low risk triggers no warning signal, only a routine follow-up reminder; medium risk triggers a yellow warning, prompting medical staff to increase monitoring frequency, such as shortening the follow-up cycle from 3 months to 1 month; high risk triggers a red warning, specifying intervention timeframes based on the level, such as a follow-up ultrasound within 72 hours for a grade 3 arteriovenous fistula and emergency intervention within 24 hours for a grade 4 arteriovenous fistula. The early warning results include risk level, warning color, intervention recommendations, and time limits. The attribution analysis results include a ranking of major risk factors, attribution weights for each factor, and an explanation of the risk transmission mechanism. These two sets of data are integrated and output in a standardized format to both healthcare worker workstations and patient terminals. Healthcare worker terminals display complete professional data and mechanism analysis, while patient terminals explain the causes of risk and precautions in plain language, such as, "Your arteriovenous fistula risk is high, mainly related to poorly controlled hypertension. Please take your medication on time and avoid lifting heavy objects with the arm on the side of the fistula." The system also supports healthcare workers entering intervention measures, automatically tracking changes in risk probability after intervention and updating early warning results, forming a closed-loop management system of early warning, intervention, and follow-up.

[0140] The aforementioned method for early warning of arteriovenous fistula (AVF) risk based on ultrasound data integrates ultrasound data with clinical historical information through multimodal data fusion and refined preprocessing, eliminating data heterogeneity and noise interference, thus laying a high-quality data foundation for risk assessment. Relying on causal discovery and effect estimation techniques, it overcomes the limitations of traditional association analysis, accurately identifying causal relationships between variables and quantifying effect strength, avoiding confounding bias and "black box" warning problems, and improving the credibility and interpretability of results. Combining counterfactual reasoning to achieve personalized risk prediction, it clarifies core risk factors and transmission mechanisms, and, in conjunction with multi-level early warning strategies, provides medical staff with targeted intervention basis, effectively compensating for the deficiencies of traditional monitoring such as insufficient personalization and delayed early warning, enhancing the scientific rigor and efficiency of clinical decision-making, promoting the transformation of AVF management towards precision and prognosticism, and providing strong technical support for the functional protection of AVF in patients.

[0141] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0142] Based on the same inventive concept, this application also provides a patient fistula risk warning system based on ultrasound data for implementing the aforementioned method for warning of patient fistula risk based on ultrasound data. The solution provided by this system is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the patient fistula risk warning system based on ultrasound data provided below can be found in the limitations of the patient fistula risk warning method based on ultrasound data described above, and will not be repeated here.

[0143] In one exemplary embodiment, such as Figure 3 As shown, a patient arteriovenous fistula risk warning system 300 based on ultrasound data is provided, including:

[0144] The data acquisition and preprocessing module 301 is used to collect patients' ultrasound data and clinical history data, and to perform multimodal preprocessing on the ultrasound data and clinical history data to obtain a standardized multimodal dataset. The ultrasound data includes structural data, hemodynamic data and examination metadata. The clinical history data includes basic information, history of underlying diseases, dialysis-related data, history of complications and medication history.

[0145] The data encoding and processing module 302 is used to classify the function of the patient's arteriovenous fistula based on the key features of B-ultrasound in the standardized multimodal dataset and through the arteriovenous fistula grading model to obtain the arteriovenous fistula function grading results and encoding features; the key features of B-ultrasound include the ratio of peak systolic velocity, the rate of increase of vessel wall thickness, and the proportion of intimal rough area.

[0146] The risk causal graph construction module 303 is used to construct a causal graph of arteriovenous fistula risk based on coding features and clinical historical data through a causal discovery algorithm, and obtain a causal graph structure representing the causal relationship between variables. The causal graph structure includes ultrasound indicators, patient baseline characteristics, treatment factors, and arteriovenous fistula functional classification as nodes.

[0147] The causal effect estimation module 304 is used to estimate the magnitude of the causal effect between variables based on the causal graph structure and the causal effect estimation method, and obtain the causal effect estimation result; the causal effect estimation result includes direct causal effect and indirect causal effect.

[0148] The risk warning and attribution analysis module 305 is used to perform personalized risk prediction based on the causal graph structure and causal effect estimation results through counterfactual reasoning, and generate risk warning results and attribution analysis results. The risk warning results include risk level and warning type, and the attribution analysis results include main risk factors and causal paths.

[0149] Furthermore, the risk cause-effect graph construction module 303 is also used for:

[0150] Based on coding features and clinical history data, the partial correlation coefficients between variables were calculated through conditional independence tests to obtain the skeleton of the undirected graph;

[0151] Based on the undirected graph skeleton, the orientation rules are applied to determine the direction of the edges by time sequence constraints, resulting in a partially directed acyclic graph.

[0152] Based on a partially directed acyclic graph, the optimal causal graph structure is obtained by optimizing the Bayesian information criterion scoring function through a greedy equivalence search algorithm.

[0153] The causal sufficiency test is performed on the optimal causal graph structure, and unobserved confounding factors are identified through residual correlation analysis to obtain the causal graph structure.

[0154] Furthermore, the risk cause-effect graph construction module 303 is also used for:

[0155] Based on the optimal cause-effect graph structure, the residual term of each node is calculated through structural equation modeling to obtain the residual sequence;

[0156] For the residual series, the correlation strength between the residuals is calculated using the Pearson correlation coefficient, resulting in the residual correlation matrix; the formula for calculating the correlation strength is:

[0157]

[0158] in, Representing variables and The residual correlation coefficient, i.e., the correlation strength, Indicates the first The variable in the first... The residual values ​​of each sample Indicates the first The mean of the residuals of each variable, Indicates the number of samples;

[0159] Based on the residual correlation matrix, significant residual correlations are identified through hypothesis testing to obtain a set of potential confounding factors;

[0160] For a set of potential confounding factors, the causal graph structure is extended using latent variable modeling methods to obtain the causal graph structure.

[0161] Furthermore, the causal effect estimation module 304 is also used for:

[0162] Based on the causal graph structure, the magnitude of direct causal effects is estimated through structural equation modeling to obtain regression coefficients;

[0163] Based on the regression coefficients, the magnitude of the causal effect is obtained by adjusting for confounding bias using a two-way stable estimation method.

[0164] For a binary treatment variable, the average treatment effect is estimated by propensity score matching to obtain the difference between the treatment group and the control group;

[0165] Based on the calibrated magnitude of the causal effect and the differences between the treatment and control groups, the causal effect estimation results are obtained.

[0166] Furthermore, based on the regression coefficients, the causal effect estimation module 304 is also used for:

[0167] Based on the causal graph structure, a propensity score is calculated for each sample using a logistic regression model, resulting in a propensity score vector. The formula for calculating the propensity score is as follows:

[0168]

[0169] in, Indicates the first The propensity score of each sample. Indicates the variable being processed. Represents a covariate vector. Represents the intercept term. Represents the coefficient vector;

[0170] Based on the propensity score vector, the sample weights are adjusted using an inverse probability weighting method to obtain a weighted sample set;

[0171] For the weighted sample set, the regression coefficients are re-estimated using the weighted least squares method to obtain the preliminary calibrated causal effect;

[0172] Based on the preliminary calibrated causal effect, the calibrated causal effect size is obtained by combining dual robust estimation with regression adjustment and propensity score weighting.

[0173] Furthermore, the risk warning and attribution analysis module 305 is also used for:

[0174] Based on the causal graph structure and the patient's current multimodal data, the posterior probability distribution of the node variables is calculated using the belief propagation algorithm to obtain the risk probability of arteriovenous fistula functional classification; the formula for calculating the risk probability is:

[0175]

[0176] in Indicates in the evidence Under certain conditions, the functional classification of arteriovenous fistula is as follows: The probability of a class Represents the normalization constant. Represents a node The potential function, Representing an edge The potential function, Indicates the number of nodes. Represents the set of evidence variables;

[0177] Based on risk probabilities, the intervention effect is simulated through counterfactual questioning, and the attribution weights of each risk factor are calculated to obtain the ranking of the main risk factors; the formula for calculating the attribution weights is:

[0178]

[0179] in, Indicates risk factors Attribution weights Indicates risk factors Conditional probability after intervention. Represents the conditional probability under the current observation conditions. This indicates the total number of risk factors. Indicates a reference value;

[0180] Based on the ranking of major risk factors, key causal chains from underlying diseases to changes in ultrasound indicators were extracted through causal path analysis, and the risk transmission mechanism was explained.

[0181] Based on risk probability, attribution weight, and risk transmission mechanism, a multi-level early warning strategy is used to generate risk early warning results and attribution analysis results. The multi-level early warning strategy triggers early warning signals of different colors according to the risk level. The risk level is divided based on the risk probability threshold.

[0182] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the patient arteriovenous fistula risk warning method based on ultrasound data as described above.

[0183] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0184] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0185] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for early warning of arteriovenous fistula risk in patients based on ultrasound data, characterized in that, The method includes: The patient's ultrasound data and clinical history data were collected, and the ultrasound data and clinical history data were preprocessed in a multimodal manner to obtain a standardized multimodal dataset. The ultrasound data included structural data, hemodynamic data and examination metadata, and the clinical history data included basic information, history of underlying diseases, dialysis-related data, history of complications and medication history. Based on the key ultrasound features in the standardized multimodal dataset, the arteriovenous fistula (AVF) function of patients is graded using an AVF grading model to obtain AVF function grading results and coding features; the key ultrasound features include the peak systolic velocity ratio, the rate of increase in vessel wall thickness, and the proportion of intimal rough areas. Based on the aforementioned coding features and clinical historical data, a causal graph of arteriovenous fistula risk is constructed using a causal discovery algorithm, resulting in a causal graph structure representing the causal relationships between variables. The causal graph structure includes ultrasound indicators, patient baseline characteristics, treatment factors, and arteriovenous fistula functional classification as nodes. Based on the causal graph structure, the magnitude of the causal effect among the variables is estimated using the causal effect estimation method, and the causal effect estimation result is obtained; the causal effect estimation result includes direct causal effect and indirect causal effect. Based on the causal graph structure and causal effect estimation results, personalized risk prediction is performed through counterfactual reasoning, generating risk warning results and attribution analysis results. The risk warning results include risk level and warning type, and the attribution analysis results include main risk factors and causal paths.

2. The method according to claim 1, characterized in that, Based on the encoded features and clinical history data, a causal graph of arteriovenous fistula risk is constructed using a causal discovery algorithm, resulting in a causal graph structure representing the causal relationships between variables, including: Based on the aforementioned coding features and clinical history data, the partial correlation coefficients between variables are calculated through conditional independence tests to obtain the skeleton of the undirected graph; Based on the undirected graph skeleton, the direction of the edges is determined by applying orientation rules through time sequence constraints to obtain a partially directed acyclic graph. Based on the aforementioned partially directed acyclic graph, the Bayesian information criterion scoring function is optimized using a greedy equivalence search algorithm to obtain the optimal causal graph structure; The optimal causal graph structure is subjected to a causal sufficiency test, and unobserved confounding factors are identified through residual correlation analysis to obtain the causal graph structure.

3. The method according to claim 2, characterized in that, The process of performing a causal sufficiency test on the optimal causal graph structure and identifying unobserved confounding factors through residual correlation analysis to obtain the causal graph structure includes: Based on the optimal cause-effect graph structure, the residual term of each node is calculated using a structural equation model to obtain the residual sequence; For the residual sequence, the correlation strength between the residuals is calculated using the Pearson correlation coefficient to obtain the residual correlation matrix; the formula for calculating the correlation strength is: in, Representing variables and The residual correlation coefficient, i.e., the correlation strength, Indicates the first The variable in the first... The residual values ​​of each sample Indicates the first The mean of the residuals of each variable, Indicates the number of samples; Based on the residual correlation matrix, significant residual correlations are identified through hypothesis testing to obtain a set of potential confounding factors; For the set of potential confounding factors, the causal graph structure is extended using latent variable modeling methods to obtain the causal graph structure.

4. The method according to claim 1, characterized in that, Based on the causal graph structure, the causal effect estimation method is used to estimate the magnitude of the causal effect among the variables, and the causal effect estimation result is obtained, including: Based on the aforementioned causal graph structure, the magnitude of direct causal effects is estimated through structural equation modeling to obtain regression coefficients; Based on the regression coefficients, the confounding bias is adjusted using a two-way stable estimation method to obtain the calibrated causal effect magnitude; For a binary treatment variable, the average treatment effect is estimated by propensity score matching to obtain the difference between the treatment group and the control group; The causal effect estimation result is obtained based on the calibrated causal effect magnitude and the difference between the treatment group and the control group.

5. The method according to claim 4, characterized in that, The step of adjusting for confounding bias using a two-way stable estimation method based on the regression coefficients to obtain the calibrated causal effect magnitude includes: Based on the aforementioned causal graph structure, a propensity score for each sample is calculated using a logistic regression model, resulting in a propensity score vector. The formula for calculating the propensity score is as follows: in, Indicates the first The propensity score of each sample. Indicates the variable being processed. Represents a covariate vector. Represents the intercept term. Represents the coefficient vector; Based on the aforementioned propensity score vector, the sample weights are adjusted using an inverse probability weighting method to obtain a weighted sample set; For the weighted sample set, the regression coefficients are re-estimated using the weighted least squares method to obtain the preliminary calibrated causal effect; Based on the preliminary causal effect, the magnitude of the calibrated causal effect is obtained by combining dual robust estimation with regression adjustment and propensity score weighting.

6. The method according to claim 1, characterized in that, Based on the causal graph structure and causal effect estimation results, personalized risk prediction is performed through counterfactual reasoning, generating risk warning results and attribution analysis results, including: Based on the causal graph structure and the patient's current multimodal data, the posterior probability distribution of the node variables is calculated using the belief propagation algorithm to obtain the risk probability of arteriovenous fistula functional classification; the formula for calculating the risk probability is: in Indicates in the evidence Under certain conditions, the functional classification of arteriovenous fistula is as follows: The probability of a class Represents the normalization constant. Represents a node The potential function, Representing an edge The potential function, Indicates the number of nodes. Represents the set of evidence variables; Based on the aforementioned risk probabilities, the attribution weights of each risk factor are calculated by simulating the intervention effect through counterfactual questioning, thus obtaining a ranking of the main risk factors. The formula for calculating the attribution weights is as follows: in, Indicates risk factors Attribution weights Indicates risk factors Conditional probability after intervention. Represents the conditional probability under the current observation conditions. This indicates the total number of risk factors. Indicates a reference value; Based on the ranking of the main risk factors, key causal chains from underlying diseases to changes in ultrasound indicators were extracted through causal path analysis, and the risk transmission mechanism was explained. Based on the risk probability, the attribution weight, and the risk transmission mechanism, a multi-level early warning strategy is used to generate risk early warning results and attribution analysis results. The multi-level early warning strategy triggers early warning signals of different colors according to the risk level. The risk level is divided based on a risk probability threshold.

7. A patient arteriovenous fistula risk early warning system based on ultrasound data, characterized in that, The system includes: The data acquisition and preprocessing module is used to collect patients' ultrasound data and clinical history data, and to perform multimodal preprocessing on the ultrasound data and clinical history data to obtain a standardized multimodal dataset. The ultrasound data includes structural data, hemodynamic data and examination metadata, and the clinical history data includes basic information, history of underlying diseases, dialysis-related data, history of complications and medication history. The data encoding and processing module is used to classify the patient's arteriovenous fistula function based on the key ultrasound features in the standardized multimodal dataset and through the arteriovenous fistula grading model, to obtain the arteriovenous fistula function grading results and encoding features; the key ultrasound features include the ratio of peak systolic velocity, the rate of increase in vessel wall thickness, and the proportion of intimal rough area. The risk causal graph construction module is used to construct a causal graph of arteriovenous fistula risk based on the encoded features and clinical historical data, and obtain a causal graph structure representing the causal relationship between variables; the causal graph structure includes ultrasound indicators, patient baseline characteristics, treatment factors and arteriovenous fistula functional classification as nodes; The causal effect estimation module is used to estimate the magnitude of the causal effect between variables based on the causal graph structure and the causal effect estimation method, and to obtain the causal effect estimation result; the causal effect estimation result includes direct causal effect and indirect causal effect. The risk warning and attribution analysis module is used to perform personalized risk prediction based on the causal graph structure and causal effect estimation results through counterfactual reasoning, and generate risk warning results and attribution analysis results; the risk warning results include risk level and warning type, and the attribution analysis results include main risk factors and causal paths.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 6.

Images