CN-121983310-A - High-risk coronary artery stenosis prediction method based on multi-level structural information collaboration

CN121983310ACN 121983310 ACN121983310 ACN 121983310ACN-121983310-A

Abstract

The invention discloses a high-risk coronary artery stenosis prediction method based on multi-hierarchy information collaboration, and belongs to the technical field of intelligent medical auxiliary decision making. Firstly, constructing a clinical pathology structure perception map to describe structure constraint relations among variables, and constructing a patient population risk conduction map to model risk association of patients in a multidimensional phenotype space. And then respectively carrying out structure perception characterization learning on the two images to obtain feature level and patient level embedded representation. And realizing cross-level semantic alignment through linear projection, introducing a learnable weight to structurally enhance the two types of representations, inputting the fused combined representation into a classifier, and outputting a risk prediction result of the high-risk coronary artery stenosis of the patient. According to the invention, through the collaborative modeling of the double graphs and the information fusion of the multi-level structure, the effective integration of the heterogeneous structure characterization is realized, and the accuracy and the stability of prediction are improved.

Inventors

SU SHUZHI
LIU SHUYI
ZHU YANMIN
LI YANTING
DAI YONG

Assignees

安徽理工大学

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (3)

1. The acute coronary syndrome high-risk vascular stenosis prediction method based on multi-level structure information cooperation is characterized by comprising the following steps of: (1) Obtaining structural clinical data of an acute coronary syndrome patient, carrying out numerical coding, missing value processing and standardization processing on the clinical data, and constructing a clinical characteristic matrix of the patient; (2) Constructing a clinical pathology structure perception graph and a patient group risk conduction graph based on the patient clinical feature matrix, wherein the clinical pathology structure perception graph is established according to the structural association relation among clinical variables and is used for describing the structural constraint and the cooperative relation of the clinical variables on the disease pathology mechanism level; (3) Firstly, respectively implementing structure perception representation learning on the clinical pathology structure perception graph and the patient group risk conduction graph to obtain a feature-level structure embedded representation and a patient-level structure embedded representation, then, processing the feature-level and patient-level embedded representation through linear structure mapping and semantic alignment mapping, optimizing the spatial distribution of the embedded representation by utilizing the linear structure mapping, ensuring the meaning consistency among different source embedded representations by utilizing the semantic alignment mapping, and finally, carrying out weighted fusion on the two types of embedded representations by utilizing a collaborative fusion module so as to generate a collaborative structure representation in a unified semantic space; (4) And inputting the collaborative structure representation into a classification prediction unit, and outputting a prediction result or occurrence probability of the high-risk coronary artery stenosis of the patient.
2. The method for predicting high-risk vascular stenosis in acute coronary syndrome based on cooperation of multi-hierarchy information according to claim 1, wherein the constructing of the clinical pathology perception map and the patient population risk conduction map in step (2) is performed by using two map models to understand the structural relationship of disease pathology layers and the potential risk association between patient populations in depth, and the steps are as follows: (2a) Clinical pathological structure perception map Constructing a clinical pathology structure perception map based on the structure association relation among clinical variables in the clinical feature matrix of the patient Wherein each feature is considered a node Corresponding to a clinical feature, the edge weights are determined based on statistical covariate characteristics of clinical variables in the patient population to obtain an adjacency matrix The method is used for expressing the structure constraint relation of clinical indexes in the occurrence and development processes of diseases; (2b) Patient population risk conductance map Constructing a patient population risk conduction diagram based on phenotype association relations among different patients in the patient clinical feature matrix Wherein each patient is a node The connection strength between the nodes is quantified by the normalized inner product of the normalized phenotype vectors, which is specifically defined as: Wherein the method comprises the steps of Representing the first in a patient population risk profile Patient and the first The strength of the connection between the patients; And Respectively represent the first Patient and the first Normalized phenotype vector for the patient; representing vectors I.e. converting the column vector into a row vector for performing a vector multiplication operation; And Respectively represent vectors And This step is used to ensure a normalization of the inner product results, resulting in a similarity score between 0 and 1; The patient graph adjacency matrix calculated by the method reflects the direction consistency of the patients in the multidimensional clinical phenotype space and is used for describing the potential risk coupling relation of the patient population level.
3. The method for predicting acute coronary syndrome high-risk vascular stenosis based on collaboration of multi-hierarchy structural information according to claim 1, wherein in the step (3), structural perception characterization learning is performed on a clinical pathology structural perception map and a patient group risk conduction map respectively, and after linear mapping and semantic alignment, the structural perception characterization learning is performed through weighting fusion by a collaborative fusion module to form a unified collaborative structural representation, and the steps are performed as follows: (3a) Structure perception feature learning of clinical pathology structure perception map Perception map for the clinical pathological structure Modeling the structural association relation between clinical variables by adopting a structural perception feature learning mechanism, weighting and aggregating node features and neighborhood structural information thereof in the graph, and inputting the processed features into a single-layer graph attention network for feature structure learning to obtain feature structure embedded representation For structural supplementation and dynamic adjustment of patient base risk characterization, wherein the weighted aggregation is specifically implemented by any node Recorded as a neighbor set On the basis, calculate it and neighbor node Attention coefficient of (a) And normalizing the weights : Wherein, the And Respectively nodes And (3) with The feature vectors after nonlinear interaction mapping, A learnable linear transformation matrix that is shared; Vector splicing operation; Is a learnable attention vector; Is usually taken In the structure perception feature learning process, nodes Adjacent node thereof Structural association weights between Is determined by the following relationship: Wherein, the And Representing nodes respectively Sum node Is used to determine the input feature vector of (a), As a learnable transformation matrix for feature space mapping, For a learnable parameter vector for characterizing node pair associations, Representing a nonlinear mapping function, and normalizing the structural association weight to obtain a node Weighting coefficients for its neighbor nodes: the weighting coefficient is used for guiding the weighted aggregation of node characteristics in the neighborhood range; (3b) Patient structure learning of patient population risk conductance map Based on the patient population risk conductance map Modeling risk association of a patient in a multidimensional clinical phenotype space, and obtaining a patient structure embedded representation by propagating and aggregating structural information of patient nodes and neighbor patient nodes thereof For characterizing the overall risk location of the patient in the population structure; (3c) Collaborative fusion mechanism Embedding representations in obtaining feature structures And patient structure embedded representation Then, introducing a collaborative fusion mechanism to realize collaborative enhancement of multi-hierarchy information, and respectively carrying out linear projection on the feature level embedding and the patient level embedding to obtain an embedded representation after projection And Wherein And (3) with For a learnable projection matrix, the method is used for mapping the embedments of different branches into a unified semantic space, and then weighting and fusing the two projected embedment representations in a collaborative fusion mode to obtain a fused final embedment representation The calculation formula is as follows: Wherein, the The initial value of the scalar weight parameter which is a leachable parameter is set to be 1.0, and the scalar weight parameter is automatically optimized through back propagation in the model training process and is used for adaptively adjusting the relative importance of the feature level information and the patient level information in the fusion process.

Description

High-risk coronary artery stenosis prediction method based on multi-level structural information collaboration Technical Field The invention relates to the technical field of artificial intelligence and medical information processing, in particular to an intelligent prediction method and system for high-risk vascular stenosis risk assessment of patients with acute coronary syndromes, and particularly relates to a high-risk coronary vascular stenosis prediction method based on multi-level structure information cooperation, which is suitable for clinical decision support systems and cardiovascular disease risk assessment scenes. Background Acute Coronary Syndrome (ACS) is an important disease type in cardiovascular diseases that leads to higher mortality and disability rates, and its central pathological basis is rupture of coronary atherosclerotic plaques and secondary thrombosis. In clinical practice, the method can identify the patients with high-risk coronary artery stenosis as early as possible, and has important guiding significance for whether emergency coronary angiography, percutaneous Coronary Intervention (PCI) and follow-up treatment strategies are implemented. The conventional risk scoring systems mainly comprise GRACE scoring, TIMI scoring and the like, are usually based on a limited number of clinical variables and are constructed in a linear weighting mode, so that complex nonlinear association relations among multidimensional clinical features are difficult to characterize, potential similarity structures among different patients cannot be reflected, and prediction accuracy and stability in complex real world groups are limited. In recent years, with the development of machine learning and deep learning technologies, some researches try to use methods such as logistic regression, support vector machines, neural networks and the like to conduct risk prediction, but most of the methods are based on sample independent isodistribution assumptions, and structural association between clinical features and similarity topological structure of a patient population level cannot be explicitly modeled, so that model generalization capability and clinical interpretation are insufficient. In addition, deep GNNs are prone to "overcomplete" problems, causing node embedment to be consistent and losing discrimination capability. Aiming at the problems, the invention provides an intelligent prediction method which can model a clinical characteristic structural relationship and a patient group similarity structure simultaneously and has a stable information fusion mechanism so as to realize accurate assessment of high-risk vascular stenosis risk of ACS patients. Disclosure of Invention Aiming at the defects in the prior art, the invention provides a high-risk coronary artery stenosis prediction method based on multi-level structure information synergy, which realizes stable fusion of multi-source structure information on the premise of not damaging the original structure specificity by collaborative modeling of a clinical pathological structure constraint relationship and a patient population risk association relationship and introducing a collaborative fusion mechanism, thereby improving prediction accuracy and robustness. The technical scheme of the invention comprises the following steps: (1) Obtaining structural clinical data of an acute coronary syndrome patient, carrying out numerical coding, missing value processing and standardization processing on the data, and constructing a clinical feature matrix of the patient: Wherein, the For the number of patient samples,Is a characteristic dimension such as biochemical index, vital sign, clinical variable and the like; (2) Multi-level structure construction The clinical pathology structure perception diagram and the patient population risk conduction diagram are constructed to deeply understand the structural relationship of disease pathology mechanism layers and potential risk association among patient populations. (2A) Clinical pathological structure perception map Constructing a clinical pathology structure perception map based on the structure association relation among clinical variables in the clinical feature matrix of the patientWherein each feature is considered a nodeFor a clinical feature, the edge weights are defined by inter-feature Pearson correlation coefficients: thereby obtaining an adjacency matrix Is used for expressing the structure constraint relation of clinical indexes in the occurrence and development processes of diseases. (2B) Patient population risk conductance map Constructing a patient population risk conduction diagram based on phenotype association relations among different patients in the patient clinical feature matrixWherein each patient is a nodeAnd calculating cosine similarity based on the standardized feature vector to obtain side weights: Wherein the method comprises the steps of Representing the first in a patient populat