Interventional surgery risk prediction method based on dual-view cross-semantic interaction

Abstract

This invention discloses a method for predicting interventional surgical risks based on dual-view cross-semantic interaction, belonging to the interdisciplinary field of artificial intelligence and intelligent diagnosis and treatment of cardiovascular diseases. Addressing the problems of feature redundancy, incomplete single-view modeling, and semantic isolation in existing methods for predicting complications during percutaneous coronary intervention, this invention's specific implementation process is as follows: First, a two-stage recursive feature elimination-extreme gradient boosting mechanism is used to screen key clinical features; second, a patient-indicator dual-view module is constructed to characterize patient similarity and indicator correlation; then, an asymmetric cross-semantic interaction module is proposed, and a collaborative inference alignment loss is designed to ensure semantic consistency, ultimately achieving dynamic information aggregation between nodes. This invention achieves accurate prediction and traceable attribution of complications during percutaneous coronary intervention, providing decision support for the early identification of high-risk patients.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and intelligent diagnosis and treatment of cardiovascular diseases. Specifically, it relates to an interventional surgery risk prediction method based on dual-view cross-semantic interaction, which is applicable to the early risk identification and warning of intraoperative complications of percutaneous coronary intervention (PCI) in patients with acute myocardial infarction (AMI). It can provide clinicians with accurate and interpretable preoperative risk assessment decision support. Background Technology

[0002] Percutaneous coronary intervention (PCI) is the core method for revascularization in patients with acute myocardial infarction (AMI), but intraoperative complications (such as no-reflow and in-stent thrombosis) are frequent and seriously threaten life. Early and accurate identification of high-risk patients is crucial for optimizing surgical strategies and improving prognosis. Current risk assessment relies on clinical scores such as GRACE, laboratory test indicators, and imaging judgments, which suffer from insufficient sensitivity, poor timeliness, and high subjectivity. In recent years, methods based on single machine learning models such as XGBoost have improved efficiency, but are still limited by indicator redundancy, weak generalization ability, and the inability to interpret the "black box" effect. Traditional graph neural networks often use single-layer isomorphic graphs, modeling only the single relationship between patients or indicators, making it difficult to integrate individual heterogeneity and the interaction between indicators; moreover, they generally use unidirectional information aggregation, ignoring the inverse sensitivity of indicators to high-risk patients and lacking medical consistency constraints. To address the aforementioned issues, this invention proposes an interventional surgery risk prediction method based on dual-view cross-semantic interaction. Through a recursive index elimination-extreme gradient boosting mechanism, a patient-index dual-view module, an asymmetric cross-semantic interaction module, and collaborative reasoning alignment loss, it achieves accurate prediction and interpretable risk attribution of intraoperative PCI complications. Summary of the Invention

[0003] This invention addresses the problems of existing methods for predicting complications during percutaneous coronary intervention (PCI) procedures, such as significant redundancy and interference from indicators, difficulty in integrating multidimensional correlations between patients and indicators using single-layer graph models, and unidirectional information aggregation mechanisms lacking clinical semantic consistency constraints. It proposes a method for predicting interventional procedure risks based on dual-view cross-semantic interaction. The specific implementation steps are as follows: 1. Recursive index elimination - extreme gradient boosting mechanism (1a) Ranking based on importance-driven metrics of XGBoost Generate an index sequence sorted in descending order of discriminative ability; its output can be represented as: in This is the complete set of preprocessed original indicators. Indicators In the XGBoost model, the importance score is defined as the average information gain brought by this metric across all decision tree split nodes. This indicates a sorting operation based on numerical values ​​from largest to smallest, and the output should be... This is a sequence of indicators arranged in descending order of importance, used to guide the subsequent recursive elimination process; (1b) Recursive index elimination and candidate subset generation By iteratively eliminating less important metrics, a subset of candidate metrics with measurable performance is constructed. This process can be formalized as follows: in To obtain the number of metrics for the current subset, from the original dimension m Decrease to the preset minimum value ,like , Indicates a sorted sequence Center front A subset consisting of indicators, for each The XGBoost model was retrained on the training set, and its AUC performance for complication prediction was computed on the independent validation set, denoted as . Ultimately, a set of "index subset-performance" pairings is obtained, providing a data foundation for the selection of the optimal subset; (1c) Redundancy filtering and final index determination by integrating dual similarity The core operation involves removing redundancy from the candidate subset with the best performance and outputting the final set of key metrics. in The subset of candidate metrics with the highest AUC on the validation set. This is a redundancy elimination function, and its execution logic is as follows: [The function is used to] eliminate redundancy. Any two indicators , Calculate linear similarity (Absolute value of Pearson correlation coefficient); Calculate nonlinear similarity ; The fusion yields the overall similarity. ; like If the pair is redundant, the pair with the higher XGBoost importance score is retained; output This is a set of final key indicators that are non-redundant, highly discriminative, and clinically significant, used to subsequently construct a patient-indicator dual-view module.

[0004] 2. Patient-Indicator Dual-View Module (2a) Construction of the patient neighborhood topology graph A patient association topology is established based on clinical phenotypic similarity, and its graph structure can be defined as an adjacency matrix: in For the total number of patients, Indicates the patient of Nearest neighbor set, which refers to the set of nearest neighbors in the standardized key indicator space, and the set of nearest neighbors with patients. Euclidean distance nearest Other patients, adjacency matrix non-zero elements This indicates that there exists a path from the patient in the patient's neighborhood topology graph. point to Undirected edges are used to encode patient similarity in clinical phenotypes; (2b) Construction of the co-dependency graph of indicators The complex relationships between explicit medical indicators are modeled, and their adjacency relationships are determined by thresholding the fusion of Pearson correlation and Gaussian kernelized Euclidean distance, which is formally represented as follows: in The number of key indicators after screening in step (1), As an indicator With indicators The fusion similarity between them This is the absolute value of the Pearson correlation coefficient between the two, used to capture linear dependence. Let be the Gaussian kernel similarity, where Indicators The vector of values ​​across all patients, and for Perform interval normalization mapping to ; To establish a preset similarity threshold, edges are only created between corresponding indicator nodes when the fusion similarity exceeds this threshold, ultimately resulting in a binary adjacency matrix for the indicator layer. ; (2c) Establishment of cross-layer association edges and construction of dual-view module The two subgraphs are coupled into a unified dual-view by using "patient-indicator" association edges. The association rule can be expressed as follows: in This is a bipartite graph adjacency matrix connecting the patient layer and the indicator layer. Indicates the patient With indicators There are cross-layer association edges, forming a dense bipartite graph connection structure between the patient layer and the indicator layer; ultimately, the dual view is defined by three parts, including the patient neighborhood topology graph. Indicator View and cross-level correlation matrix ,in and These are patient node sets and indicator node sets, respectively, which realize the integrated representation of the network linking individual patient information and indicators.

[0005] 3. Asymmetric cross-semantic interaction module (3a) Calculation of patient attention to indicators Each patient can adaptively focus on the clinical indicators most relevant to their risk, and their cross-attention output can be expressed as: in The patient node is projected and embedded as a query. The indicator node is projected and embedded as a key. The value is embedded as the projection of the indicator node. For embedded dimensions, This is a scaling factor used to stabilize the gradient; output This represents the contextual information that patients obtain from the aggregated indicator layer, enabling patients to focus on high-risk indicators in a personalized way; (3b) Indicators to calculate patient attention Each clinical indicator can inversely identify a high-risk patient subgroup that is sensitive to it, and its cross-attention output can be expressed as: in The indicator node is projected and embedded as a query. The patient node projection is embedded as a key. Project the patient node as a value; output This indicates that the indicators are aggregated from the patient level to obtain contextual information, enabling the indicators to perceive their discriminative contribution in specific high-risk populations.

[0006] (3c) Asymmetric bidirectional information aggregation and node representation update By fusing the cross-attention outputs from both directions, the node representation update mechanism is defined as follows: in It is a non-linear activation function. To output the projection matrix, For hidden layer dimensions; patient branch output enhances patient representation. Directly used for complication risk classification tasks; the indicator branches are updated to represent the indicators through residual connections. It participates in consistency constraints as an auxiliary semantic anchor.

[0007] 4. Collaborative Inference Alignment Loss (4a) Define cross-modal semantic consistency constraints Construct a regularization term to measure the semantic alignment between patients and indicator representations, with the following mathematical form: in and The first Projection embedding and fusion representation of each patient node and The first The projected embedding and updated representation of each indicator node; this loss term measures the alignment between the projected embedding and the interactively enhanced representation, providing an auxiliary consistency constraint signal for the asymmetric structure.

[0008] (4b) Constructing the Collaborative Inference Alignment Loss Function Combining the above consistency constraint with the standard classification task loss, we form the final optimization objective, whose expression is: in The weighted cross-entropy classification loss is defined as: in For the first The true complication label for each patient The probability of complications predicted by the model; The collaborative reasoning alignment loss defined in step (4a), To balance the hyperparameters, the weight of the consistency constraint in the total loss is adjusted.

[0009] (4c) Implementing a two-way calibration mechanism for semantic alignment By minimizing the collaborative inference alignment loss function, collaborative optimization of the patient and indicator representation spaces is achieved during training; in implementation, it can be... and As an alignment target, and by stopping its gradient propagation when computing the consistency term, the consistency constraint mainly acts on the projection branch, thereby guiding the projection embedding alignment interaction to enhance semantics in the form of soft regularization.

[0010] The method of the present invention has the following advantages: (1) This invention adopts a recursive index elimination-extreme gradient boosting mechanism, combined with redundant control to automatically extract key clinical predictive factors that are most discriminative for complications of percutaneous coronary intervention, effectively reducing high-dimensional noise interference and enhancing the model's generalization ability. (2) This invention innovatively constructs a patient-indicator dual-view module. The patient neighborhood topology graph depicts the association topology between patients based on clinical phenotype similarity, while the indicator co-dependency graph explicitly models the complex coupling relationship between medical indicators by integrating linear and nonlinear measures; (3) By designing an asymmetric cross-semantic interaction module, this invention achieves deep and dynamic interaction between patient nodes and indicator nodes. Patients can adaptively focus on the key indicators that contribute the most to their own risk, while the indicators can also identify high-risk patient subgroups that are sensitive to them; (4) The present invention constructs a collaborative reasoning alignment loss, which optimizes classification performance while forcing the patient representation space and the indicator representation space to maintain semantic consistency. Attached Figure Description

[0011] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0012] Figure 1 This is the overall framework diagram of the present invention. Figure 2 This is a classification result diagram based on a confusion matrix on a binary classification dataset. Detailed Implementation

[0013] The specific implementation steps of this invention are as follows: 1. Data Preprocessing Data Collection: Clinical data were collected from 811 patients with acute myocardial infarction (AMI) before and after percutaneous coronary intervention (PCI). The data included 51 raw clinical indicators.

[0014] Data imputation: Missing values ​​are imputed using the mean (numerical) or mode (categorical); outliers are corrected based on clinical expert advice; categorical variables are converted to numerical form using one-hot encoding.

[0015] Dataset standardization: The Z-Score standardization method is used to process numerical indicators to eliminate the influence of units. Let the mean of a certain indicator be... The standard deviation is The standardized formula is as follows: 2. Recursive metric elimination - extreme gradient boosting mechanism Indicator Importance Assessment: Train an XGBoost classifier on the training set and use its built-in Gain metric to calculate the importance score for each indicator, which represents the average information gain brought by the indicator across all decision tree split nodes.

[0016] Recursive index elimination: Sort all indices in descending order of importance score, starting from the full index set, iteratively remove the indices with the lowest importance, and retrain the XGBoost model with the remaining indices after each removal, and evaluate its AUC performance on the validation set.

[0017] Optimal subset selection: Iterate through all possible subset sizes and select the subset with the highest AUC on the validation set as the first candidate key metric set. .

[0018] Redundancy filtering: for Any two indicators and Calculate its linear correlation (Absolute value of Pearson correlation coefficient) and nonlinear similarity Set fusion weights The overall similarity is .like If a pair is found to be redundant, it is considered a redundant pair, and the pair with the higher XGBoost importance score is retained. The final output is a set of key metrics without redundancy. .

[0019] 3. Patient-Indicator Dual-View Module Patient neighborhood topology construction: Each patient is considered a node, with its initial node index being... Using the standardized value vector on the index space, the K nearest neighbor algorithm is used to find the K other patients whose Euclidean distance is closest to each patient in the index space, and undirected edges are established between them to form a patient similarity graph.

[0020] Metric Co-dependency Graph Construction: Treat each key metric as a node, and calculate the comprehensive similarity between any two metrics only if... At that time, weighted edges are established between the corresponding indicator nodes, and the weights are... Construct an indicator correlation network.

[0021] Cross-layer connections: via bipartite graph adjacency matrix Connecting the patient layer and the indicator layer, where elements Indicates the first The patient in the Standardized values ​​for each indicator (missing values ​​are set to 0).

[0022] 4. Asymmetric cross-semantic interaction module Patient to cross-attention indicators Embed the patient projection As a query, indicator projection embedding As keys and values, the cross-attention output is obtained after scaling the dot product and applying softmax. This allows patients to adaptively monitor high-risk indicators.

[0023] Indicators to patient cross-attention Embedding indicator projection As a query, patient projection embedding As keys and values, the cross-attention output is obtained after scaling the dot product and applying softmax. This enables the indicators to provide a reverse perception of sensitive patient groups.

[0024] Information aggregation: Cross-attention output Embedded with patient projection Concatenation, through learnable fusion functions Patient characteristics after fusion And used for risk classification; at the same time, Cross-attention output With the output of the index encoder By adding the residuals, we can obtain It serves as an auxiliary semantic anchor point in consistency constraints.

[0025] Information aggregation: Cross-attention output Embedded with patient projection splicing, through learnable linear transformation With nonlinear activation functions Patient characteristics after fusion And used for risk classification; at the same time, Cross-attention output With feature encoder output By adding the residuals, we can obtain It serves as an auxiliary semantic anchor point in consistency constraints.

[0026] 5. Collaborative Inference Alignment Loss Classification loss: Weighted cross-entropy loss is used. The primary objective is to optimize the model's accuracy in predicting complications.

[0027] Define the alignment loss for collaborative inference: Loss function: , where hyperparameters This is used to balance classification accuracy and semantic consistency. Minimize during training. This ensures that the representations learned by the model are both accurate and consistent with medical logic.

[0028] The effectiveness of this invention was further verified through the following experiments: To fully verify the effectiveness and superiority of the interventional surgery risk prediction method based on dual-view cross-semantic interaction proposed in this invention, a rigorous experimental evaluation was conducted on real-world clinical cohort data from the South Branch of Shanghai Sixth People's Hospital. This cohort included 811 patients with acute myocardial infarction (AMI) undergoing percutaneous coronary intervention (PCI), encompassing 51 multidimensional clinical indicators, covering demographics, medical history, clinical test indicators, imaging parameters, and surgical records. A stratified sampling strategy was employed, dividing the entire dataset into training, validation, and test sets in a 6:2:2 ratio to ensure a consistent complication rate distribution across subsets. During the training phase, the training set data underwent standardized preprocessing, and a recursive indicator elimination-extreme gradient boosting mechanism was applied to select key indicator subsets for modeling. Subsequently, based on the selected indicators, a dual-view module was constructed on the training set: the patient neighborhood topology graph uses the K-nearest neighbor algorithm to construct similarity connections between patients based on Euclidean distance; the indicator collaborative dependency graph calculates the association strength between indicators by fusing Pearson correlation coefficient and Gaussian kernelized Euclidean distance, and forms connections after thresholding. Based on this dual-view structure, the model achieves dynamic information interaction between patient nodes and indicator nodes through an asymmetric cross-semantic interaction module, and integrates collaborative inference alignment loss for end-to-end optimization. Finally, the trained model was evaluated on a completely independent test set to objectively measure the predictive ability of this invention for intraoperative PCI complications.

[0029] This invention used 811 patients with acute myocardial infarction (AMI) who met the fourth edition of the global unified definition of myocardial infarction. All patients underwent emergency coronary angiography and were diagnosed with culprit vascular stenosis within 12 hours of symptom onset. Or acute occlusion with clear indications for PCI. The ultimate prediction goal is whether serious intraoperative complications (IOCs) occur during PCI, including coronary artery perforation, in-stent thrombosis, major hemorrhage, and adverse events such as slow flow / no-reflow. This outcome label was independently determined by two professionally trained cardiovascular clinical researchers, and cross-checking was used to ensure accuracy. A total of 51 clinical indicators were input, and after screening by the recursive indicator elimination-extreme gradient enhancement mechanism of this invention, a concise and non-redundant subset of key indicators was obtained for modeling. All data collection, validation, and submission were completed by a professional clinical team to ensure the authenticity and reliability of the data. The above experiments not only demonstrated the superior performance of this invention in high-precision risk prediction, but also provided a solid technical foundation for achieving accurate and reliable intraoperative risk stratification in clinical practice.

Claims

1. A method for predicting interventional surgical risks based on dual-view cross-semantic interaction, characterized in that, The method includes the following steps: (1) An index screening mechanism based on recursive index elimination and extreme gradient boosting is adopted. The extreme gradient boosting classifier is used as the base learner for recursive index elimination to obtain a set of candidate key indicators on the training set. Furthermore, the redundancy of the candidate set is controlled based on Pearson correlation to obtain the final set of key predictive factors. (2) A patient-index dual-view module is constructed. By modeling the phenotypic similarity between patients and the association structure between clinical indicators, the dual heterogeneity of individual patient differences and the dependence between indicators is explicitly characterized. (3) An asymmetric cross-semantic interaction module is designed. A patient-centered asymmetric bidirectional cross-attention mechanism is adopted to realize the information exchange of heterogeneous semantic space. The patient node is the main query. The index semantic information is aggregated through cross-attention and fused through learnable projection to directly serve risk prediction. The index node focuses on the patient representation through a lightweight residual path as an auxiliary semantic anchor. (4) Design a collaborative reasoning alignment loss to measure the alignment degree between bibranch projection embedding and interactive enhancement representation, provide auxiliary consistency constraint signals for asymmetric structures, and achieve accurate prediction of complications during percutaneous coronary intervention.

2. The interventional surgery risk prediction method based on dual-view cross-semantic interaction according to claim 1, characterized in that... The patient-indicator dual-view module described in step (2) explicitly characterizes the dual heterogeneity of individual patient differences and the dependencies between indicators by separately modeling the association structure between phenotypic similarities and clinical indicators among patients. Its overall structure is defined as follows: in For the patient node, Total number of patients; As an indicator node, These are the filtered indicator dimensions. This is a binary adjacency matrix of the patient's neighborhood topology graph. This is a binary adjacency matrix for the index co-dependency graph; This represents the patient-indicator cross-layer correlation matrix; where the patient neighborhood topology graph is located. Characterizing phenotypic similarity among patients, indicator co-dependency plot Modeling the relationship structure between indicators, where and These are the patient node set and the indicator node set, respectively, and the cross-layer correlation matrix. Connect the two subgraphs to form a unified dual-view structure; (2a) The patient neighborhood topology graph is constructed based on the set of key predictive factors obtained in step (1) to build phenotypic similarity associations between patients. In the standardized key indicator space, the neighborhood connection relationship is determined according to the indicator vector distance between patients, and the nearest neighbor is selected for each patient. Undirected edges are established for each of the other patients, forming a sparse patient similarity topology with an adjacency matrix. elements Indicates the patient With patients They exhibit high similarity in clinical phenotype and have established connections. The neighborhood size hyperparameter controls the sparsity of the graph; (2b) The aforementioned co-dependency graph explicitly models the complex relationships between medical indicators, and its binary adjacency matrix is ​​defined as: in Edge connection threshold; fusion similarity Determined by fusion of linear correlation and Gaussian kernelized Euclidean distance: in To integrate weights, As an indicator With indicators The absolute value of the Pearson correlation coefficient; To achieve Gaussian kernel similarity based on the differences in indicator vectors, a nonlinear similarity between indicators is calculated using a Gaussian kernel function, where... For kernel function hyperparameters, and for Perform interval normalization mapping to ; (2c) The cross-layer association matrix couples the two subgraphs into a unified dual view through patient-indicator association edges, and its elements are defined as follows: This indicates the establishment of a dense, fully connected relational structure between the patient layer and the indicator layer, enabling the integrated representation of individual patient information and the indicator relational network.

3. The interventional surgery risk prediction method based on dual-view cross-semantic interaction according to claim 1, characterized in that... The design of the asymmetric cross-semantic interaction module described in step (3) employs a patient-centered asymmetric bidirectional cross-attention mechanism to achieve information exchange in heterogeneous semantic spaces. The steps are as follows: the asymmetric bidirectional information aggregation and representation update mechanism is defined as follows: in Embed a linearly projected representation of the patient. This is the raw output of the index encoder. It is a non-linear activation function. To output the projection matrix, For the hidden layer dimension; To achieve patient-indicator cross-attention output, a scaling dot product attention mechanism is used to enable adaptive patient attention to high-risk indicators. To achieve cross-attention output from indicators to patients, a scaling dot product attention mechanism is used to enable the reverse perception of indicators for sensitive patient groups. This module represents the vector concatenation operation; its asymmetric design is specifically reflected in: (3a) The patient branch adopts a splicing-projection fusion mechanism: embedding the patient projection Attention output aggregated from the indicator layer The data is spliced ​​together and then subjected to a learnable linear transformation. With nonlinear activation functions Generate enhanced patient representation It is directly used for downstream risk classification tasks; (3b) The index branch adopts a lightweight residual connection mechanism: the original output of the index encoder is used to... Attention output aggregated from the patient level The residuals are summed to generate the updated index representation. , as an auxiliary semantic anchor point to participate in consistency regularization; (3c) This asymmetric design reflects a patient-centered prediction orientation: the patient branch undertakes the main classification learning task, and the index branch provides auxiliary regularization signals through consistency constraints.

4. The interventional surgery risk prediction method based on dual-view cross-semantic interaction according to claim 1, characterized in that... The collaborative reasoning alignment loss described in step (4) measures the alignment degree between the bi-branch projection embedding and the interaction-enhanced representation, providing an auxiliary consistency constraint signal for the asymmetric structure. Its constraint objective is defined as: in and The first Projection embedding and fusion representation of each patient node and The first The projection embedding and updated representation of each indicator node; the patient consistency constraint measures the alignment between the patient projection embedding and the fused representation; the indicator consistency constraint measures the alignment between the indicator projection embedding and the updated representation. In terms of implementation, it can be and As an alignment objective, it stops gradient propagation, making the consistency constraint primarily apply to the projection branch; The overall optimization objective is: in For the classification loss term, a weighted cross-entropy approach is used to drive the model to learn the correct complication prediction labels. To balance the hyperparameters, the weight of the consistency constraint in the total loss is adjusted.

Images