Intelligent judgment method for stenosis grading based on real-time monitoring of vascular access fistula flow

Abstract

This invention discloses an intelligent method for grading stenosis based on real-time monitoring of arteriovenous fistula (AVF) flow, belonging to the field of medical monitoring technology. This invention collects and preprocesses instantaneous blood flow, arterial pressure, and venous pressure signals from the AVF to calculate the arteriovenous pressure difference and blood flow impedance; extracts time-series data to construct a three-dimensional multi-dimensional time-series sequence sample; employs an improved BiLSTM network to mine the positive trend and lag correlation features of the time series, and constructs a fusion feature set by combining statistical features; constructs an ensemble regression model based on gradient boosting regression tree, random forest regression, and support vector regression, optimizes the weights of each model through K-fold cross-validation, and outputs a weighted fusion of continuous stenosis quantitative scores; generates adaptive segmented thresholds by combining K-means clustering with clinical diagnostic criteria, mapping the quantitative scores to classify stenosis levels. This invention achieves automated, intelligent, and accurate grading of AVF stenosis, possessing profound application value.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, specifically to an intelligent method for determining the stenosis grading based on real-time monitoring of arteriovenous fistula flow in vascular access. Background Technology

[0002] With the number of patients with end-stage renal disease increasing year by year, hemodialysis has become a core treatment for maintaining patients' lives. Arteriovenous fistulas (AVFs) are the preferred vascular access method in clinical practice due to their low infection rate and long lifespan. AVF stenosis, as the most common complication, leads to insufficient blood flow and inadequate dialysis, and in severe cases, AVF occlusion, forcing patients to rebuild the access and significantly increasing treatment costs and risks. Currently, AVF stenosis monitoring technology is rapidly developing, evolving from traditional manual palpation and auscultation to ultrasound, Doppler blood flow detection, and real-time monitoring with wearable devices. These technologies are widely used in clinical settings such as nephrology and hemodialysis centers, providing crucial support for maintaining AVF function. According to clinical research statistics, the incidence of AVF stenosis among hemodialysis patients in my country exceeds 40%, and early, accurate monitoring can improve AVF patency by more than 30%. However, existing monitoring technologies still have certain shortcomings. Traditional manual monitoring is highly subjective and has low accuracy, making it difficult to achieve real-time continuous monitoring; imaging examinations such as ultrasound have a lag and cannot capture early, minute lesions; some intelligent monitoring methods rely solely on blood flow or pressure parameters, failing to consider individual physiological differences and temporal dynamic characteristics, resulting in insufficient accuracy in stenosis grading. In this context, several related studies have been conducted within the field, such as:

[0003] Chinese patent (publication number CN112998743A) discloses a method, system and wearable medical device for assessing the degree of arteriovenous fistula stenosis. The method assesses the degree of arteriovenous fistula stenosis by collecting murmur signals and extracting spectral centroid features and auditory spectral flux features. However, the method relies only on murmur signal features and does not combine hemodynamic core parameters. It is easily affected by external environmental noise and lacks dynamic analysis capabilities, making it difficult to accurately reflect the dynamic changes in stenosis.

[0004] Chinese patent (publication number CN111134652A) discloses a multifunctional monitoring wristband for arteriovenous fistulas in hemodialysis patients. This method monitors parameters such as the natural blood flow, vessel wall thickness, and limb temperature of the fistula to achieve real-time monitoring and abnormal alarm of the fistula status. However, the method for collecting and processing multiple parameters is relatively simple and cannot achieve quantitative grading of the degree of stenosis, making it difficult to meet the needs of accurate clinical judgment.

[0005] Chinese patent (publication number CN117815476A) discloses a stenosis monitoring device after arteriovenous fistula repair surgery. This method collects fistula signals through vibration, sound and pressure sensors, automatically detects changes in fistula thrill and murmur, and realizes auxiliary monitoring of postoperative stenosis. However, this method does not perform deep feature extraction on the collected signals, judges stenosis by simple signal feature thresholds, the grading standard is relatively crude, and does not consider the influence of individual physiological differences of patients on the monitoring results, resulting in limited accuracy.

[0006] In summary, most existing solutions only focus on parameter acquisition and simple analysis, without combining deep learning and ensemble regression algorithms to achieve precise quantification and intelligent grading of the degree of narrowness, making it difficult to meet the actual needs of early, accurate, and continuous clinical monitoring. Summary of the Invention

[0007] To address the aforementioned technical issues, this application discloses an intelligent method for determining stenosis grading based on real-time monitoring of arteriovenous fistula flow, specifically including:

[0008] Instantaneous blood flow, arterial pressure, and venous pressure of the vascular access fistula were collected. After preprocessing the collected signals, the corrected arteriovenous pressure difference and corrected blood flow impedance were calculated.

[0009] A sliding window is constructed based on time steps, and multidimensional time series samples are constructed based on instantaneous blood flow, arteriovenous pressure difference, and blood flow impedance.

[0010] A time-series deep feature representation model is constructed using an improved bidirectional long short-term memory network (BiLSTM). The model takes multi-dimensional time-series samples as input and outputs time-series deep features. Combined with the statistical features of the collected data, a fusion feature set is constructed.

[0011] An ensemble regression model is constructed using multiple regression algorithms. The input is a fused feature set, and the output is a predicted continuous narrow quantized score.

[0012] Based on the continuous stenosis quantification score, combined with the preset adaptive segmentation threshold, the continuous stenosis quantification score is mapped to classify the levels, and the vascular access fistula stenosis classification result is obtained.

[0013] Preferably, the calculation of the corrected arteriovenous pressure difference and blood flow impedance is specifically based on: the collected arterial end pressure and venous end pressure, combined with the time decay coefficient. and individual physiological correction factors The dynamically corrected arteriovenous pressure gradient is calculated using the following formula:

[0014]

[0015] in, This refers to the arteriovenous pressure difference. Arterial end pressure, For venous end pressure, As an individual physiological correction factor, The time decay coefficient, The update formula is: ,in, This is the updated time decay factor. The preset update rate coefficient, , These are the instantaneous blood flow at the current moment and the previous moment, respectively.

[0016] Preferably, the corrected blood flow impedance is specifically based on the dynamically corrected arteriovenous pressure difference. With instantaneous blood flow Based on the patient's real-time body temperature and the elasticity coefficient of the arteriovenous fistula, the corrected blood flow impedance is calculated using the following formula:

[0017]

[0018]

[0019] in, This is the corrected blood flow impedance. This is a blood flow viscosity correction term. Normal human core body temperature The baseline coefficient of blood viscosity at that time, The preset temperature influence coefficient, Provide the patient's real-time temperature. The elastic coefficient of the arteriovenous fistula is fitted from the baseline value of the vessel diameter.

[0020] Preferably, the multidimensional time-series sequence sample specifically comprises: a preset sliding window with a length of [missing information]. The sliding step size is Based on the collected instantaneous blood flow Arteriovenous pressure difference Blood flow impedance The signal data within each sliding window is extracted in chronological order to construct a multidimensional time series sample.

[0021] Preferably, the time-series deep feature representation model specifically involves: inputting multi-dimensional time-series samples, capturing the dynamic changes of instantaneous blood flow, arteriovenous pressure difference, and blood flow impedance over time using a forward LSTM of an improved bidirectional long short-term memory network (BiLSTM), extracting retrospective features and lag correlation features using a backward LSTM, concatenating the forward and backward time-series features to obtain the original output features of the BiLSTM, and calculating the weights of the bidirectional output features through an attention mechanism layer to obtain a deep feature set.

[0022] Preferably, the improved bidirectional long short-term memory network (BiLSTM) specifically includes an input layer, a BiLSTM hidden layer, an attention mechanism layer, and an output layer. In the input layer, temporal attention gating is used to assign weights to the three-dimensional features at each time step, strengthening the representation of key time-step features related to arteriovenous fistula stenosis. The formula is as follows:

[0023]

[0024] in, For enhanced input features, For the temporal attention gating matrix, For the original input, For gating bias terms, For element-wise multiplication, For softmax normalization;

[0025] A hierarchical dropout layer is added after the output of each BiLSTM hidden layer. The dropout probability is adaptively adjusted according to the number of layers, as shown in the formula:

[0026]

[0027]

[0028] in, For the first Dropout probability of hidden layers As the baseline dropout probability, Adjust the rate for dropout; The hidden layer outputs the feature vector after hierarchical dropout processing. For dropout probability, This is the original feature vector output by the hidden layer;

[0029] At the attention mechanism layer, the spliced ​​features output from the BiLSTM hidden layer are processed through a fully connected layer. Mapping to feature vectors of the same dimension, calculating attention weights, and assigning key feature weights to the original output features of BiLSTM yields preliminary temporal deep features, as shown in the formula:

[0030]

[0031]

[0032] in, For attention weights, These are the spliced ​​features output from the BiLSTM hidden layer. , These are the weight matrix and bias term of the fully connected layer, respectively. , These are the weight matrix and bias term of the attention layer, respectively; For preliminary time-series deep features, , The first Attention weights and splicing features at each time step;

[0033] Redundant features are suppressed in the output layer using an L1 regularization term, as shown in the formula:

[0034]

[0035] in, The deep features output by the model. , These are the fusion weight matrix and the fusion bias term, respectively. The regularization coefficient is . For feature dimension, For the fusion weight matrix The Each weighted element.

[0036] Preferably, the integrated regression model is constructed based on three regression models: Gradient Boosting Regression Tree (GBRT), Random Forest Regression (RFR), and Support Vector Regression (SVR). It adopts an improved integrated architecture that combines a weighted fusion strategy with cross-validation optimization to achieve accurate prediction of narrow quantization scores and output the predicted continuous narrow quantization scores.

[0037] Preferably, the three regression models are: input fused feature vector The three basic regression models, GBRT, RFR, and SVR, were trained independently.

[0038] The GBRT model employs a gradient boosting iterative strategy, using a decision tree as the base learner. In each iteration, a weak decision tree is trained, and the model parameters are updated by minimizing the mean squared error (MSE) between the predicted value and the true narrow score.

[0039] The RFR model is trained in parallel by constructing multiple decision trees. Each decision tree is trained based on randomly sampled samples and features, and the output is the average of the prediction results of each decision tree. During the training process, the model is optimized by limiting the depth of the decision trees and the number of leaf nodes.

[0040] The SVR model uses the radial basis function (RBF) as the kernel function. By optimizing the penalty coefficient C and the kernel function parameter γ, it fits the nonlinear mapping relationship between the feature vector V and the true narrowing score.

[0041] Preferably, the improved integrated architecture specifically employs K-fold cross-validation to optimize the weight coefficients of the three basic models. The optimization objective is to minimize the mean square error between the predicted value and the actual narrowing score, as shown in the formula:

[0042]

[0043] in, For the first The weights of each model, For the first The mean squared error of cross-validation for each model. The mean squared error of the cross-validation of the three basic models;

[0044] The prediction results of the three basic models are weighted and summed to obtain the narrow quantitative score of the prediction of the integrated regression model.

[0045] Preferably, the adaptive segmentation threshold is specifically as follows: collect data of clinically diagnosed arteriovenous fistula stenosis cases, extract the actual stenosis degree and corresponding continuous stenosis quantitative score sample set for each case, use K-means clustering algorithm combined with clinical diagnostic criteria, set the number of clusters for K-means clustering according to clinical stenosis grading criteria, obtain cluster centers, and use the midpoint between cluster centers as the adaptive segmentation threshold to calculate the threshold.

[0046] Compared with the prior art, the technical solution of this application has the following technical effects:

[0047] This invention collects multi-source physiological signals of instantaneous blood flow, arterial pressure, and venous pressure in vascular access fistulas, and introduces time decay coefficient, individual physiological correction factor, real-time body temperature, and vascular elasticity coefficient for dynamic correction. It simultaneously considers individual physiological differences and time decay interference, effectively improving the accuracy and adaptability of hemodynamic parameter calculation.

[0048] This invention employs an improved bidirectional long short-term memory network architecture to construct a temporal deep feature representation model. It mines temporal positive change trends and backtracking lag correlation features respectively, and combines attention mechanism and regularization constraints to complete deep feature purification. This results in stronger feature representation capabilities and better resistance to noise and redundancy interference.

[0049] This invention integrates multiple classic regression models to construct an ensemble regression architecture. By constructing an ensemble regression model and adaptively allocating the weights of each model through cross-validation, it achieves accurate output of narrow quantitative scores and improves the reliability of narrow quantitative assessment.

[0050] This invention uses clinical cases combined with clustering algorithms to generate adaptive segmented thresholds to classify stenosis levels, which aligns with clinical diagnostic standards. The grading boundaries are more reasonable and adaptable, enabling automated, standardized, and intelligent grading of arteriovenous fistula stenosis and meeting the application needs of real-time clinical monitoring and early screening.

[0051] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0052] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0054] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0055] Figure 1 The flowchart shows the overall process of intelligent stenosis classification determination method based on real-time monitoring of arteriovenous fistula flow.

[0056] Figure 2 This is an architecture diagram of an intelligent method for determining stenosis grading based on real-time monitoring of arteriovenous fistula flow.

[0057] Figure 3 This is an architecture diagram of the temporal deep feature representation model in this application;

[0058] Figure 4 This is a diagram of the architecture of the integrated regression model in this application;

[0059] Figure 5 This is a data diagram of the training process of each method in the embodiments of this application;

[0060] Figure 6 This is a comparison chart of the quantitative analysis results of the stenosis degree of each method in the embodiments of this application on typical mild stenosis samples;

[0061] Figure 7 This is a comparison chart of the accuracy probability density of each method in the embodiments of this application for each type of narrow sample;

[0062] Figure 8 This is a comparison chart of the overall performance data of the methods in the embodiments of this application. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0064] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0065] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0066] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0067] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0068] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0069] Example 1 mainly describes a method for intelligent determination of stenosis grading based on real-time monitoring of arteriovenous fistula flow in vascular access, such as... Figure 1 , Figure 2 As shown, it specifically includes:

[0070] Instantaneous blood flow, arterial pressure, and venous pressure of the vascular access fistula were collected. After preprocessing the collected signals, the corrected arteriovenous pressure difference and corrected blood flow impedance were calculated.

[0071] A sliding window is constructed based on time steps, and multidimensional time series samples are constructed based on instantaneous blood flow, arteriovenous pressure difference, and blood flow impedance.

[0072] A time-series deep feature representation model is constructed using an improved bidirectional long short-term memory network (BiLSTM). The model takes multi-dimensional time-series samples as input and outputs time-series deep features. Combined with the statistical features of the collected data, a fusion feature set is constructed.

[0073] An ensemble regression model is constructed using multiple regression algorithms. The input is a fused feature set, and the output is a predicted continuous narrow quantized score.

[0074] Based on the continuous stenosis quantification score, combined with the preset adaptive segmentation threshold, the continuous stenosis quantification score is mapped to classify the levels, and the vascular access fistula stenosis classification result is obtained.

[0075] Furthermore, preprocessing includes denoising and outlier removal. Denoising employs a wavelet thresholding algorithm, the process of which is as follows: The db4 wavelet is selected as the base wavelet, and the acquired raw signal is decomposed into three layers of wavelets to obtain the level 1, level 2, and level 3 detail coefficients d1, d2, and d3, and the level 3 approximation coefficient a3. The noise standard deviation is calculated using a noise estimation formula to determine the wavelet threshold. The formula is:

[0076]

[0077]

[0078] in, The standard deviation of noise. For wavelet threshold, This represents the number of sampling points for the original signal.

[0079] Thresholding is applied to the detail coefficients d1, d2, and d3 using the following formula:

[0080]

[0081] in, The detail coefficients after processing. For the first Layer detail factor, A sign function to represent positive and negative signs;

[0082] Outlier removal is performed by replacing outliers using linear interpolation of adjacent data.

[0083] Furthermore, the corrected arteriovenous pressure difference and blood flow impedance are calculated, specifically based on the collected arterial and venous pressures, combined with the time decay coefficient. ( The value is adaptively adjusted based on the blood flow fluctuation frequency (range 0.92-0.98) and an individual physiological correction factor. The dynamically corrected arteriovenous pressure gradient is calculated using the following formula:

[0084]

[0085] in, This refers to the arteriovenous pressure difference. Arterial end pressure, For venous end pressure, Individual physiological correction factor (APC) obtained by fitting the patient's baseline weight and blood pressure values. =0.01×body weight + 0.85×baseline blood pressure), The time decay coefficient, The update formula is: ,in, This is the updated time decay factor. The preset update rate coefficient, , These are the instantaneous blood flow at the current moment and the previous moment, respectively.

[0086] Furthermore, the corrected blood flow impedance is specifically based on the dynamically corrected arteriovenous pressure difference. With instantaneous blood flow Based on the patient's real-time body temperature and the elasticity coefficient of the arteriovenous fistula, the corrected blood flow impedance is calculated using the following formula:

[0087]

[0088]

[0089] in, This is the corrected blood flow impedance. This is a blood flow viscosity correction term. Normal human core body temperature The baseline coefficient of blood viscosity at that time, The preset temperature influence coefficient, Provide the patient's real-time temperature. The elastic coefficient of the arteriovenous fistula is fitted from the baseline value of the vessel diameter.

[0090] Furthermore, the multidimensional time-series sequence samples specifically include: a preset sliding window with a length of [missing information]. The sliding step size is ( The value range is 5-15 sampling periods. The value ranges from 1 to 3 sampling periods; based on the collected instantaneous blood flow. Arteriovenous pressure difference Blood flow impedance The signal data within each sliding window is extracted in chronological order to construct a multidimensional time series sample.

[0091] Furthermore, such as Figure 3 The diagram shows the architecture of the temporal deep feature representation model. Specifically, the temporal deep feature representation model is as follows: inputting multi-dimensional temporal sequence samples, the forward LSTM of the improved bidirectional long short-term memory network BiLSTM captures the dynamic changes of instantaneous blood flow, arteriovenous pressure difference, and blood flow impedance over time, and the backward LSTM extracts retrospective features and lag correlation features. The forward and backward temporal features are concatenated to obtain the original output features of BiLSTM, and the deep feature set is obtained by weighting the bidirectional output features through the attention mechanism layer.

[0092] Furthermore, the improved Bidirectional Long Short-Term Memory (BiLSTM) network specifically includes an input layer, a BiLSTM hidden layer, an attention mechanism layer, and an output layer. In the input layer, temporal attention gating is used to weight the 3D features at each time step, strengthening the representation of key time-step features related to arteriovenous fistula stenosis. The formula is as follows:

[0093]

[0094] in, For enhanced input features, For the temporal attention gating matrix, For the original input, For gating bias terms, For element-wise multiplication, For softmax normalization;

[0095] A hierarchical dropout layer is added after the output of each BiLSTM hidden layer. The dropout probability is adaptively adjusted according to the number of layers, as shown in the formula:

[0096]

[0097]

[0098] in, For the first Dropout probability of hidden layers As the baseline dropout probability, Adjust the rate for dropout; The hidden layer outputs the feature vector after hierarchical dropout processing. For dropout probability, This is the original feature vector output by the hidden layer;

[0099] At the attention mechanism layer, the spliced ​​features output from the BiLSTM hidden layer are processed through a fully connected layer. Mapping to feature vectors of the same dimension, calculating attention weights, and assigning key feature weights to the original output features of BiLSTM yields preliminary temporal deep features, as shown in the formula:

[0100]

[0101]

[0102] in, For attention weights, These are the spliced ​​features output from the BiLSTM hidden layer. , These are the weight matrix and bias term of the fully connected layer, respectively. , These are the weight matrix and bias term of the attention layer, respectively; For preliminary time-series deep features, , The first Attention weights and splicing features at each time step;

[0103] Redundant features are suppressed in the output layer using an L1 regularization term, as shown in the formula:

[0104]

[0105] in, The deep features output by the model. , These are the fusion weight matrix and the fusion bias term, respectively. The regularization coefficient is . For feature dimension, For the fusion weight matrix The Each weighted element.

[0106] Furthermore, the input layer of the improved bidirectional long short-term memory network BiLSTM is configured with a temporal attention gating structure to receive three-dimensional feature input. The baseline dropout probability is set to 0.1, the adjustment rate is set to 0.1, the gating matrix is ​​3 rows and 3 columns, and the bias term is a fixed value. The enhanced input feature dimension is consistent with the original input dimension.

[0107] The BiLSTM hidden layer has a multi-layer structure with a baseline dropout probability of 0.1 and an adjustment rate of 0.1. Each layer has a dropout layer, the number of neurons in the hidden layer is 64, and the concatenated feature dimension is T×64 (T is the sliding window length). The output feature after dropout processing has the same dimension as the original hidden layer output feature.

[0108] The attention mechanism layer is a fully connected layer with 32 neurons. It receives the features concatenated from the hidden layer and outputs attention weights. After weight allocation, preliminary deep features are obtained. The number of parameters in this layer matches the dimension of the output features of the hidden layer.

[0109] The output layer receives the initial features from the attention mechanism layer, configures the feature fusion structure, sets the regularization coefficient to 0.01, and ensures that the feature dimension is consistent with the input feature dimension. It outputs a deep feature vector with the output dimension consistent with the input feature dimension.

[0110] Furthermore, the statistical characteristics of the collected data specifically include the statistical characteristics of instantaneous blood flow, arteriovenous pressure difference, and blood flow impedance, as well as the correlation characteristics among the three.

[0111] Among them, the statistical characteristics of a single parameter include mean, standard deviation, coefficient of variation, maximum value, minimum value, peak-to-peak value, skewness, and kurtosis;

[0112] The correlation between the three factors is calculated using the Pearson correlation coefficient, and the formula is as follows:

[0113]

[0114] in, The Pearson correlation coefficient between two characteristic parameters. , The corresponding features are the first. 1 eigenvalue, for The sampling mean of the feature parameter. for The sampling mean of the feature parameter. This represents the total number of sampling points.

[0115] Furthermore, such as Figure 4 The diagram shows the architecture of the ensemble regression model. Specifically, the ensemble regression model is constructed based on three regression models: Gradient Boosting Regression Tree (GBRT), Random Forest Regression (RFR), and Support Vector Regression (SVR). It adopts an improved ensemble architecture that combines a weighted fusion strategy with cross-validation optimization to achieve accurate prediction of narrow quantization scores and output the predicted continuous narrow quantization scores.

[0116] Furthermore, the three regression models are as follows: input fused feature vector The three basic regression models, GBRT, RFR, and SVR, were trained independently.

[0117] The GBRT model employs a gradient boosting iterative strategy, using decision trees as base learners. Each iteration trains a weak decision tree, and the model parameters are updated by minimizing the mean squared error (MSE) between the predicted and true narrow scores. The prediction formula is as follows:

[0118]

[0119] in, This represents the predicted output of the GBRT model, where M is the number of base learners. For the first The predicted output of a weak decision tree. To fuse feature vectors;

[0120] The RFR model is trained in parallel by constructing multiple decision trees. Each decision tree is trained based on randomly sampled samples and features, and the output is the mean of the prediction results of each decision tree. During training, the model is optimized by limiting the depth of the decision trees (ranging from 5 to 10 layers) and the number of leaf nodes (ranging from 10 to 20). The prediction formula is as follows:

[0121]

[0122] in, This is the predicted output of the RFR model. For the number of decision trees, For the first The predicted output of each decision tree;

[0123] The SVR model uses a radial basis function (RBF) as the kernel function. The penalty coefficient C (ranging from 0.1 to 10) and the kernel parameter γ (ranging from 0.001 to 0.01) are determined through optimization. A nonlinear mapping relationship between the feature vector V and the true narrowing score is fitted, and the prediction formula is as follows:

[0124]

[0125] in, This is the predicted output of the SVR model. For support vector coefficients, The true narrow score corresponding to the support vectors. For radial basis kernel functions, For support vectors, This is a bias term.

[0126] Furthermore, the Gradient Boosting Regression Tree (GBRT) adopts a structure combining basic decision trees with gradient iteration. It consists of multiple basic decision trees connected in series, with each basic decision tree serving as an iteration unit. By iteratively updating the model parameters step by step, a complete gradient boosting regression structure is formed. The number of basic decision trees is set to 100, the depth of each decision tree is controlled at 3-5 layers, the iteration step size is set to 0.01, and the regularization coefficient is set to 0.001. During model training, the parameters are adaptively optimized to ensure fitting accuracy.

[0127] Furthermore, Random Forest Regression (RFR) consists of multiple independent decision trees operating in parallel. Each decision tree is independent of the others and has no hierarchical nesting, forming a parallel computation structure. It does not require gradient iteration and directly outputs the combined prediction results of multiple decision trees. The number of decision trees is set to 50, the depth of each decision tree is controlled at 4-6 layers, the random sampling ratio is set to 0.8, the feature sampling ratio is set to 0.7, the regularization coefficient is set to 0.001, and the learnable parameters are adaptively adjusted through model training.

[0128] Furthermore, Support Vector Regression (SVR) adopts a simple structure of input feature mapping, kernel function mapping, and output regression, without complex hierarchical nesting. It only includes three basic modules: feature input, kernel function mapping, and regression output. Its kernel function is a radial basis function with a parameter of 0.01. The feature mapping dimension is consistent with the input feature dimension, the regularization coefficient is set to 0.001, and the output regression result matches the input feature dimension. The learnable parameters are adaptively optimized through model training.

[0129] Furthermore, the improved integrated architecture specifically employs K-fold cross-validation to optimize the weight coefficients of the three base models. The optimization objective is to minimize the mean squared error between the predicted value and the actual narrowing score, as shown in the formula:

[0130]

[0131] in, For the first The weights of each model, For the first The mean squared error of cross-validation for each model. The mean squared error of the cross-validation of the three basic models;

[0132] The prediction results of the three basic models are weighted and summed to obtain the narrow quantitative score of the prediction of the integrated regression model.

[0133] Furthermore, the adaptive segmentation threshold is specifically implemented as follows: Data on clinically diagnosed arteriovenous fistula stenosis cases are collected. The true degree of stenosis and the corresponding continuous stenosis quantitative score sample set for each case are extracted. A K-means clustering algorithm combined with clinical diagnostic criteria is used. Based on the clinical stenosis grading standards (usually divided into mild, moderate, severe, and occlusion levels), the number of clusters in the K-means clustering is set to cluster the sample set, obtaining cluster centers. The midpoint between the cluster centers is used as the adaptive segmentation threshold, and the threshold is calculated using the following formula:

[0134]

[0135] in, For the first A threshold, For the first Cluster centers.

[0136] Furthermore, taking grade four as an example, the quantitative scores of continuous stenosis are mapped to classify the grades, resulting in the grading of arteriovenous fistula stenosis. The specific mapping rules are as follows:

[0137] When continuous narrow quantization score At that time, it was determined to be mild stenosis, and the stenosis classification was Grade I; when At that time, it was determined to be moderate stenosis, and the stenosis grade was II; At that time, it was determined to be severe stenosis, and the stenosis grade was III; At that time, it was determined to be an arteriovenous fistula occlusion, with a stenosis grade of IV;

[0138] in, The predicted values ​​output by the integrated regression model.

[0139] This embodiment details an intelligent method for grading stenosis based on real-time monitoring of arteriovenous fistula (AVF) flow. It collects and preprocesses instantaneous blood flow, arterial pressure, and venous pressure signals from the AVF to calculate the arteriovenous pressure difference and blood flow impedance. Time-series data is extracted to construct a three-dimensional multi-dimensional time-series sequence sample. An improved BiLSTM network is used to mine the positive trend and lag correlation features of the time series, and a fusion feature set is constructed by combining statistical features. An ensemble regression model is built based on gradient boosting regression trees, random forest regression, and support vector regression. The weights of each model are optimized through K-fold cross-validation, and a weighted fusion output of a continuous stenosis quantitative score is obtained. An adaptive segmentation threshold is generated by combining K-means clustering with clinical diagnostic criteria, and the quantitative score is mapped to classify the stenosis level.

[0140] Example 2, based on Example 1, details a comparative experiment on the analysis of arteriovenous fistula stenosis characteristics using the publicly available MIT-BIH hemodynamic monitoring dataset and existing methods. The specific process is as follows:

[0141] The PhysioNet-MIT-BIH-Hemodynamic-Monitoring-Dataset (MIT-BIH Hemodynamic Monitoring Public Dataset) from the PhysioNet public physiological signal resource platform was used as the dataset for this experiment. The dataset was divided into training, validation, and test sets in a 7:2:1 ratio. The test set contained a total of 488 multidimensional time-series samples of arteriovenous fistulas, including 122 normal samples, 122 mild stenosis samples, 122 moderate stenosis samples, and 122 severe stenosis samples.

[0142] The comparison methods used were existing classic models such as LSTM-NFS (Long Short-Term Memory Network Analysis), BiLSTM-IFS (Traditional Bidirectional Long Short-Term Memory Network Analysis), CNN-LSTM-ISG (Convolutional-Long Short-Term Memory Hierarchical Method), GBDT-IFQ (Gradient Boosting Decision Tree Quantization Method), and XGBoost-ISS (Extreme Gradient Boosting Regression Scoring Method).

[0143] Among them, LSTM-NFS (Long Short-Term Memory Network Analysis) filters and preprocesses the original signal of the dataset, extracts instantaneous blood flow, arterial pressure, and venous pressure time series data to construct samples, inputs them into a standard LSTM network to mine time series features, and outputs stenosis classification results.

[0144] BiLSTM-IFS (Traditional Bidirectional Long Short-Term Memory Network Analysis) denoises the collected hemodynamic parameters, constructs multidimensional time-series samples, extracts forward and backward time-series features using a traditional BiLSTM network and splices them together, and outputs the arteriovenous fistula stenosis monitoring results based on the features.

[0145] CNN-LSTM-ISG (Convolutional-Long Short-Term Memory Grading Method) preprocesses the original temporal signal, extracts local features of the signal through CNN, and then inputs them into the LSTM network to mine temporal correlation features. After fusing the two types of features, the grading of arteriovenous fistula stenosis is completed.

[0146] GBDT-IFQ (Gradient Boosting Decision Tree Quantization) preprocesses instantaneous blood flow and arteriovenous pressure difference data, extracts statistical features to construct a feature set, trains the model using the GBDT algorithm, outputs the quantification score of arteriovenous fistula stenosis, and completes the classification.

[0147] XGBoost-ISS (Extreme Gradient Boosting Regression Scoring Method) preprocesses the physiological parameters of the vascular access, constructs a feature vector, trains the model through the XGBoost regression algorithm, outputs the arteriovenous fistula stenosis score, and classifies the stenosis level according to a fixed threshold.

[0148] The implementation process of this method is as follows: the signal sampling frequency is preset to 10Hz, matching the original sampling specifications of the public dataset; the temporal sliding window length is set to 12, and the sliding step size is set to 3; the initial baseline value of the time decay coefficient is set to 0.92, the update rate coefficient is set to 0.08, and the baseline value of the individual physiological correction factor is set to 1.05; the baseline value of the blood flow viscosity coefficient is set to 0.96, the temperature influence coefficient is set to 0.025, and the baseline value of the arteriovenous fistula elasticity coefficient is set to 1.12; the baseline dropout probability of the improved bidirectional long short-term memory network is set to 0.1, the dropout adjustment rate is set to 0.1, and the L1 regularization coefficient is set to 0.005; the ensemble regression model adopts a 5-fold cross-validation setting, and the K-means clustering algorithm presets the number of clusters to 4 according to the clinical four-level stenosis grading standard.

[0149] This method and its models were pre-trained using the training set, achieving a convergence accuracy of 96.86% after 69 rounds of training. The pre-training data for each method was statistically analyzed, resulting in Table 1 below, which shows the training process data for each method.

[0150] Table 1. Training process data for each method

[0151] Indicator Type LSTM-NFS BiLSTM-IFS CNN-LSTM-ISG GBDT-IFQ XGBoost-ISS This method Convergence cycles 64 rounds 70 rounds 86 rounds 93 rounds 65 rounds 69 rounds Training convergence accuracy 85.37% 91.12% 88.45% 93.26% 89.93% 96.86%

[0152] According to Table 1 and Figure 5 As shown in the training process data graphs, the training convergence accuracy of each comparative method is distributed in the range of 85.37% to 93.26%. Among them, the GBDT-IFQ method has the highest training convergence accuracy, but it also has the most convergence rounds. In contrast, the training convergence accuracy of this method reaches 96.86%, which is higher than all other comparative models. Moreover, the number of convergence rounds is 69, which is relatively low among all models. The training iteration rhythm is stable and reasonable, without premature convergence or late iteration, while achieving a high degree of fit.

[0153] After the model pre-training is completed, the test machine data is preprocessed by filtering and denoising. Based on the baseline parameters of pressure, flow and body temperature in the dataset, time decay coefficient, individual physiological correction factor, real-time body temperature and vascular elasticity coefficient are introduced to complete the dynamic correction of arteriovenous pressure difference and blood flow impedance respectively. The sliding window length is set to 12 and the sliding step size is 3. The three-dimensional time series data within the window is extracted according to the time step to construct multi-dimensional time series sequence samples.

[0154] The constructed multidimensional time-series samples are input into the improved bidirectional long short-term memory network of this invention. The forward LSTM is used to capture the dynamic change trend of hemodynamic parameters over time, and the backward LSTM extracts time-series backtracking features and lag correlation features. After time-series attention-gated weight allocation, hierarchical adaptive Dropout regularization, attention mechanism layer feature weighting, and output layer L1 redundancy suppression, the deep time-series features are output. A fusion feature set is constructed by combining the sample statistical features.

[0155] An ensemble regression architecture was built based on gradient boosting regression tree, random forest regression, and support vector regression. K-fold cross-validation was used to optimize the weight coefficients of each basic model, and the weighted fusion was used to output the continuous stenosis quantitative score. Based on the stenosis level samples labeled in the dataset, the K-means clustering algorithm was used to determine the adaptive segmentation threshold to complete the intelligent classification of arteriovenous fistula stenosis levels.

[0156] This method and various comparative methods were used to analyze 488 samples in the test set, and the classification accuracy data for different types of narrowing samples are shown in Table 2 below:

[0157] Table 2. Grading accuracy data for samples of different stenosis types.

[0158] Indicator Type LSTM-NFS BiLSTM-IFS CNN-LSTM-ISG GBDT-IFQ XGBoost-ISS This method Normal sample accuracy 87.45% 90.32% 91.06% 92.18% 91.53% 96.37% Accuracy of mildly narrow samples 82.14% 85.67% 86.29% 87.05% 86.42% 92.58% Accuracy of moderately narrow samples 84.36% 87.51% 88.17% 88.93% 88.26% 94.15% Accuracy of severely narrow samples 86.72% 89.84% 90.53% 91.47% 90.89% 95.62% Overall accuracy 85.18% 88.35% 89.07% 90.02% 89.46% 94.69%

[0159] According to Table 1 and Figure 6 The comparison chart of the quantitative analysis results of the stenosis degree of the various methods on typical mild stenosis samples shows that the time series analysis curve of this method is highly consistent with the actual stenosis quantitative score change trend. It can more accurately and stably depict the dynamic law of mild stenosis state evolution over time. Compared with other comparative models, it has better accuracy and robustness in time series quantitative fitting and dynamic feature capture.

[0160] According to Table 2 and Figure 7 As shown in the probability density comparison chart of the accuracy of each method for each type of stenosis sample, the probability density peak of this method is more concentrated, the dispersion range is smaller, and the degree of multi-peak shift is lower. It can maintain a higher and more stable accuracy distribution under all four types of stenosis samples. Compared with the other comparison methods, it has better recognition accuracy and robustness, and has stronger adaptive characterization and stable recognition ability for vascular stenosis samples of different degrees.

[0161] The overall performance data for each method was statistically analyzed, resulting in the comprehensive performance table in Table 3 below:

[0162] Table 3 Comprehensive Performance Table

[0163] Indicator Type LSTM-NFS BiLSTM-IFS CNN-LSTM-ISG GBDT-IFQ XGBoost-ISS This method Overall accuracy 85.18% 88.35% 89.07% 90.02% 89.46% 94.69% Time consumed per sample analysis 18.65ms 21.32ms 24.58ms 12.36ms 13.75ms 15.42ms Mean Absolute Error 0.19 0.16 0.14 0.13 0.14 0.07 Narrow grading accuracy 84.82% 87.65% 88.39% 89.27% 88.74% 94.31% Mild narrowing recognition accuracy 82.14% 85.67% 86.29% 87.05% 86.42% 92.58% Feature extraction effectiveness 78.35% 81.46% 83.72% 85.18% 84.53% 91.67%

[0164] In Table 4, the overall accuracy is based on all 486 time-series samples in the test set, and the corresponding value is obtained by statistically analyzing the percentage of correctly identified samples out of the total samples. The time taken for a single sample analysis is the average computation time taken for 100 repeated inferences on a single sample under a unified hardware and software environment. The accuracy of narrowing classification is the percentage of correctly identified samples in each type of narrowing sample out of the total number of narrowing samples. The accuracy of mild narrowing identification is based on 121 mild narrowing samples in the test set, and the result is obtained by statistically analyzing the percentage of correctly identified samples out of the total number of samples in this type. The feature extraction effectiveness is the normalized score of the combined effective feature ratio of the statistical model and the feature extraction time, which is converted to a percentage to obtain the feature extraction efficiency value.

[0165] According to Table 4 and Figure 8 As shown in the comparison chart of the comprehensive performance data of the various methods, this method has significant advantages in all core evaluation indicators. Its comprehensive performance is superior to the existing comparative models. It can effectively improve the accuracy, stability and feature mining ability of vascular stenosis identification and classification while ensuring real-time operation.

[0166] This embodiment details a comparative experiment on the vascular access arteriovenous fistula stenosis feature analysis using the publicly available MIT-BIH hemodynamic monitoring dataset and existing methods. During the process, this invention mines the temporal positive change trend and retrospective lag correlation features through a time-series deep feature characterization model, achieves accurate output of stenosis quantification scores through an integrated regression model, and completes the stenosis level classification through adaptive segmented thresholds. This realizes the automated, standardized, and intelligent classification of vascular access arteriovenous fistula stenosis, and its comprehensive performance is significantly better than existing typical methods, proving that this method can meet the application needs of clinical real-time monitoring and early screening.

[0167] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for intelligent determination of stenosis grading based on real-time monitoring of arteriovenous fistula flow, characterized in that, include: Instantaneous blood flow, arterial pressure, and venous pressure of the vascular access fistula were collected. After preprocessing the collected signals, the corrected arteriovenous pressure difference and corrected blood flow impedance were calculated. A sliding window is constructed based on time steps, and multidimensional time series samples are constructed based on instantaneous blood flow, arteriovenous pressure difference, and blood flow impedance. A time-series deep feature representation model is constructed using an improved bidirectional long short-term memory network (BiLSTM). The model takes multi-dimensional time-series samples as input and outputs time-series deep features. Combined with the statistical features of the collected data, a fusion feature set is constructed. An ensemble regression model is constructed using multiple regression algorithms. The input is a fused feature set, and the output is a predicted continuous narrow quantized score. Based on the continuous stenosis quantification score, combined with the preset adaptive segmentation threshold, the continuous stenosis quantification score is mapped to classify the levels, and the vascular access fistula stenosis classification result is obtained.

2. The intelligent stenosis classification method based on real-time monitoring of arteriovenous fistula flow according to claim 1, characterized in that, The calculation of the corrected arteriovenous pressure difference and blood flow impedance is specifically based on: the collected arterial end pressure and venous end pressure, combined with the time decay coefficient. and individual physiological correction factors The dynamically corrected arteriovenous pressure gradient is calculated using the following formula: ， in, This refers to the arteriovenous pressure difference. Arterial end pressure, For venous end pressure, As an individual physiological correction factor, The time decay coefficient, The update formula is: ,in, This is the updated time decay factor. The preset update rate coefficient, , These are the instantaneous blood flow at the current moment and the previous moment, respectively.

3. The intelligent stenosis grading method based on real-time monitoring of arteriovenous fistula flow according to claim 2, characterized in that, The corrected blood flow impedance is specifically based on the dynamically corrected arteriovenous pressure difference. With instantaneous blood flow Based on the patient's real-time body temperature and the elasticity coefficient of the arteriovenous fistula, the corrected blood flow impedance is calculated using the following formula: ， ， in, This is the corrected blood flow impedance. This is a blood flow viscosity correction term. Normal human core body temperature The baseline coefficient of blood viscosity at that time, The preset temperature influence coefficient, Provide the patient's real-time temperature. The elastic coefficient of the arteriovenous fistula is fitted from the baseline value of the vessel diameter.

4. The intelligent method for determining stenosis grading based on real-time monitoring of arteriovenous fistula flow according to claim 1, characterized in that, The multidimensional time-series sequence sample specifically comprises: a preset sliding window with a length of [missing information]. The sliding step size is Based on the collected instantaneous blood flow Arteriovenous pressure difference Blood flow impedance The signal data within each sliding window is extracted in chronological order to construct a multidimensional time series sample.

5. The intelligent method for determining stenosis grading based on real-time monitoring of arteriovenous fistula flow according to claim 1, characterized in that, The aforementioned temporal deep feature representation model is as follows: inputting multi-dimensional temporal sequence samples, using a forward LSTM of an improved bidirectional long short-term memory network (BiLSTM) to capture the dynamic changes of instantaneous blood flow, arteriovenous pressure difference, and blood flow impedance over time, using a backward LSTM to extract retrospective features and lag correlation features, concatenating the forward and backward temporal features to obtain the original output features of the BiLSTM, and using an attention mechanism layer to calculate the weights of the bidirectional output features to obtain a deep feature set.

6. The intelligent stenosis grading method based on real-time monitoring of arteriovenous fistula flow according to claim 5, characterized in that, The improved bidirectional long short-term memory network (BiLSTM) specifically includes an input layer, a BiLSTM hidden layer, an attention mechanism layer, and an output layer. In the input layer, temporal attention gating is used to assign weights to the three-dimensional features at each time step, strengthening the representation of key time-step features related to arteriovenous fistula stenosis. The formula is as follows: ， in, For enhanced input features, For the temporal attention gating matrix, For the original input, For gating bias terms, For element-wise multiplication, For softmax normalization; A hierarchical dropout layer is added after the output of each BiLSTM hidden layer. The dropout probability is adaptively adjusted according to the number of layers, as shown in the formula: ， ， in, For the first Dropout probability of hidden layers As the baseline dropout probability, Adjust the rate for dropout; The hidden layer outputs the feature vector after hierarchical dropout processing. For dropout probability, This is the original feature vector output by the hidden layer; At the attention mechanism layer, the spliced ​​features output from the BiLSTM hidden layer are processed through a fully connected layer. Mapping to feature vectors of the same dimension, calculating attention weights, and assigning key feature weights to the original output features of BiLSTM yields preliminary temporal deep features, as shown in the formula: ， ， in, For attention weights, These are the spliced ​​features output from the BiLSTM hidden layer. , These are the weight matrix and bias term of the fully connected layer, respectively. , These are the weight matrix and bias term of the attention layer, respectively; For preliminary time-series deep features, , The first Attention weights and splicing features at each time step; Redundant features are suppressed in the output layer using an L1 regularization term, as shown in the formula: ， in, The deep features output by the model. , These are the fusion weight matrix and the fusion bias term, respectively. The regularization coefficient is . For feature dimension, For the fusion weight matrix The Each weighted element.

7. The intelligent method for determining stenosis grading based on real-time monitoring of arteriovenous fistula flow according to claim 1, characterized in that, The ensemble regression model is specifically constructed based on three regression models: Gradient Boosting Regression Tree (GBRT), Random Forest Regression (RFR), and Support Vector Regression (SVR). It adopts an improved ensemble architecture that combines a weighted fusion strategy with cross-validation optimization to achieve accurate prediction of narrow quantization scores and output the predicted continuous narrow quantization scores.

8. The intelligent method for determining stenosis grading based on real-time monitoring of arteriovenous fistula flow according to claim 7, characterized in that, The three regression models are specifically: input fused feature vector The three basic regression models, GBRT, RFR, and SVR, were trained independently. The GBRT model employs a gradient boosting iterative strategy, using a decision tree as the base learner. In each iteration, a weak decision tree is trained, and the model parameters are updated by minimizing the mean squared error (MSE) between the predicted value and the true narrow score. The RFR model is trained in parallel by constructing multiple decision trees. Each decision tree is trained based on randomly sampled samples and features, and the output is the average of the prediction results of each decision tree. During the training process, the model is optimized by limiting the depth of the decision trees and the number of leaf nodes. The SVR model uses the radial basis function (RBF) as the kernel function. By optimizing the penalty coefficient C and the kernel function parameter γ, it fits the nonlinear mapping relationship between the feature vector V and the true narrowing score.

9. The intelligent method for determining stenosis grading based on real-time monitoring of arteriovenous fistula flow according to claim 8, characterized in that, The improved integrated architecture specifically employs K-fold cross-validation to optimize the weight coefficients of the three base models. The optimization objective is to minimize the mean square error between the predicted value and the actual narrowing score, as shown in the formula: ， in, For the first The weights of each model, For the first The mean squared error of cross-validation for each model. The mean squared error of the cross-validation of the three basic models; The prediction results of the three basic models are weighted and summed to obtain the narrow quantitative score of the prediction of the integrated regression model.

10. The intelligent method for determining stenosis grading based on real-time monitoring of arteriovenous fistula flow according to claim 1, characterized in that, The adaptive segmentation threshold is specifically calculated as follows: Collect data on clinically diagnosed arteriovenous fistula stenosis cases, extract the actual degree of stenosis and the corresponding continuous stenosis quantitative score sample set for each case, use the K-means clustering algorithm combined with clinical diagnostic criteria, set the number of clusters for K-means clustering according to the clinical stenosis grading criteria, obtain cluster centers, and use the midpoint between the cluster centers as the adaptive segmentation threshold to calculate the threshold.

Images