Method and system for constructing a risk assessment model for patients with myocardial infarction

Abstract

The present invention discloses a method and system for constructing a risk assessment model for myocardial infarction patients, specifically relating to the field of cardiovascular disease prediction, and includes the following steps: I. Collecting data from various modalities of hospital information systems, image archiving systems, and wearable devices, and integrating data from different modalities. The present invention can optimize the accuracy of myocardial infarction risk prediction, so that risk assessment not only relies on statistical characteristics, but also integrates pathophysiological mechanisms, improves the clinical interpretability of predictions, avoids the limitations of traditional data-driven methods that are only based on correlation modeling, can dynamically adjust the best treatment plan, realize data-driven personalized medicine, and improve the scientific nature of intervention decisions; can quickly adapt to new patient data, improve individualized prediction capabilities, improve data utilization, enhance the model's generalization ability for different patient groups, avoid training the entire model from scratch, significantly reduce computing resource consumption, make up for the problem of insufficient data for a single patient, and improve the prediction ability of rare cases.

Description

Technical Field

[0001] The present invention relates to the field of cardiovascular disease prediction, and in particular to a method and system for constructing a risk assessment model for myocardial infarction patients. Background Art

[0002] Myocardial infarction is one of the major cardiovascular diseases that cause death and disability worldwide, and its morbidity and mortality rates are increasing year by year. Since the occurrence of myocardial infarction is affected by multiple factors, including physiological indicators, hemodynamic characteristics, and lifestyle factors, how to accurately assess the risk of myocardial infarction in individual patients has become an important topic in clinical diagnosis and preventive medicine. Traditional risk assessment methods are usually based on statistical models or simple machine learning methods, which fail to fully utilize the potential information of multimodal data and make it difficult to achieve personalized and dynamic risk prediction. In addition, current models often lack causal reasoning capabilities, making it difficult to effectively evaluate the impact of different interventions on the risk of myocardial infarction. Therefore, there is an urgent need for a risk assessment model for myocardial infarction patients that integrates multimodal data, has causal reasoning and personalized adaptability, to improve prediction accuracy and provide a scientific basis for personalized medicine and early intervention.

[0003] After searching, Chinese patent number CN119324059A discloses a method and system for constructing a risk assessment model for myocardial infarction patients. Although this invention makes the assessment model construction technology more perfect by optimizing the construction technology of the risk assessment model for myocardial infarction patients, it only relies on statistical characteristics for risk assessment, which reduces the clinical interpretability of the prediction. It has the limitation of modeling based only on correlation, and cannot dynamically adjust the optimal treatment plan, which reduces the scientific nature of intervention decisions. In addition, the existing method and system for constructing a risk assessment model for myocardial infarction patients cannot quickly adapt to new patient data, which reduces the individualized prediction ability. The model has poor generalization ability for different patient groups. The entire model needs to be retrained for new patients, which increases computing resource consumption, and the prediction ability for rare cases is low. Therefore, we propose a method and system for constructing a risk assessment model for myocardial infarction patients. Summary of the Invention

[0004] The purpose of the present invention is to solve the above-mentioned problems and propose a method and system for constructing a risk assessment model for myocardial infarction patients.

[0005] In a first aspect of the present invention, a method and system for constructing a risk assessment model for myocardial infarction patients are first proposed. The method comprises:

[0006] Ⅰ. Collect data from hospital information systems, imaging archiving systems, and wearable devices, and integrate data from different modalities;

[0007] II. Extract patient characteristic data, analyze the impact of hemodynamic parameters on the occurrence of myocardial infarction, and identify key factors affecting the risk of myocardial infarction;

[0008] III. Train a personalized risk assessment model and build a lightweight edge risk prediction model based on real-time data streams to perform personalized risk prediction for different patients;

[0009] IV. Detect potentially abnormal samples, compare risk assessment results across different patient groups, analyze the model's adaptability in different scenarios, and continuously optimize model parameters based on clinical expert feedback.

[0010] As a further solution of the present invention, the specific steps of integrating different modal data in step I are as follows:

[0011] S1.1: Z-score and normalize the collected ECG, coronary artery CT scan, and laboratory test data to shrink each modality to a preset data interval. Linear interpolation is used to align the time series information in each modality, and the sliding window method is used to synchronize the sampling frequency of each modality.

[0012] S1.2: Traverse each modal data and process the missing values ​​in each modal data by filling the mean. Then check whether each data deviates from the mean by three times the standard deviation. If it deviates, the corresponding data is judged as an outlier and is removed or replaced with the mean. Then, the wavelet transform is used to remove the noise information in each modal data.

[0013] S1.3: Use CNN and LSTM encoding networks to extract the special information of different modal data respectively and map them to the same feature space to obtain the unimodal features of each modal data in the same feature space. Based on the contribution of different modalities, different weights are assigned to each unimodal feature. Then, all unimodal features are weighted and combined to form a global feature vector.

[0014] As a further solution of the present invention, the specific calculation formula for Z-score standardization described in S1.1 is as follows:

[0015] ;

[0016] ;

[0017] Where, represents the mean of the data; represents the total number of data; Representative Original data values; represents the standard deviation of the data; Representative The standardized data values;

[0018] The specific calculation formula for the normalization process described in S1.1 is as follows: ;

[0019] Where, Representative Normalized data values; Representative Original data values; Represents the maximum value in multimodal data; Represents the minimum value in multimodal data.

[0020] As a further embodiment of the present invention, the specific steps of analyzing the effect of hemodynamic parameters on the occurrence of myocardial infarction in step II are as follows:

[0021] S2.1: Collect the generated global feature vectors to construct a multimodal feature set ,in Indicates the The feature vector of time steps, and , by using linear transformation to map each input data in the multimodal feature to a high-dimensional space to obtain the input embedding vector, where;

[0022] S2.2: Calculate the query, key, and value matrices for each input embedding vector. Based on the query and key matrices, calculate the self-attention weights between the input embedding vectors at different time steps through different attention heads. Normalize the attention scores using the Softmax function, and calculate the weighted sum of the attention of each input embedding vector. Then, concatenate the weighted sum results of each attention head to generate the output features.

[0023] S2.3: Perform nonlinear transformation on the output features of each time step through the FFN network, perform layer normalization on the output features after nonlinear transformation, add the original global feature vector to the output features to generate the corresponding high-dimensional feature matrix, and then obtain the vascular structure information through the coronary artery CTA image data, and set the vascular centerline as , the vessel radius is set to ,in is the arc length coordinate along the centerline of the blood vessel;

[0024] S2.4: Construct a three-dimensional geometric model of the coronary arteries based on the acquired vascular structure information and divide the computational grid. Establish a mathematical model of blood flow in the coronary arteries using the Navier-Stokes equations. Set boundary conditions for the coronary artery inlet and outlet, as well as the vessel wall, based on the established blood flow mathematical model.

[0025] S2.5: By solving the blood flow mathematical model, the blood flow velocity field at different locations within the blood vessel is obtained, and the maximum flow velocity in the stenotic region of the vessel is calculated. Based on the blood flow velocity field at different locations and the three-dimensional geometric model of the coronary artery, the wall shear stress and coronary artery stenosis rate are calculated to obtain fluid dynamics simulation data for each coronary artery;

[0026] S2.6: Perform weighted fusion on the generated high-dimensional feature matrix and coronary artery fluid dynamics simulation data to form a global feature representation. Establish a myocardial infarction risk assessment model through logistic regression. Use this model to calculate the probability of myocardial infarction. Then, based on the probability of myocardial infarction for each hemodynamic parameter, calculate the characteristic contribution of different hemodynamic parameters in the model. Analyze the impact of the numerical range of hemodynamic parameters on the occurrence of myocardial infarction, and calculate the critical value for high risk of myocardial infarction.

[0027] S2.7: Use the SHAP method to visualize the probability of myocardial infarction, the characteristic contribution of different hemodynamic parameters in the model, and the critical value of high risk of myocardial infarction, demonstrate the influence of each hemodynamic parameter on the prediction of myocardial infarction, give the patient's comprehensive risk score, and mark high-risk factors.

[0028] As a further solution of the present invention, the specific calculation formula for the linear change in S2.1 is as follows: Where, Represents the time step Input feature embedding of Represents a trainable weight matrix; Represents the time step The input data, ; represents the bias term;

[0029] The specific calculation formula for the weighted sum of the input embedding vector attention described in S2.2 is as follows: ;

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] Where, Represents input feature embedding; represents the query matrix; represents the bond matrix; Representative value matrix; 、 as well as Represent a set of trainable weight matrices respectively; Represents the attention score matrix, which is used to compare the correlation between features at different time steps; ; represents the dimension of the bond matrix; Represents the feature matrix after self-attention calculation; represents the normalization function; Represents multi-head attention output; Representative The result of an attention head; represents the sum function; represents the linear transformation parameter;

[0035] The specific expression of the blood flow mathematical model described in S2.4 is as follows: Where, represents blood density; represents the blood flow velocity vector; Represents the current time; represents the velocity gradient; Represents blood pressure; represents blood dynamic viscosity; Laplace operator representing velocity; since coronary blood flow is pulsating, the momentum equation needs to be combined with the continuity equation , to ensure; the specific boundary conditions of the blood flow mathematical model are as follows: ;

[0036] in, Representative Moment coronary artery inlet flow velocity; represents the average flow velocity; represents the pulsation amplitude; Represents the basic frequency corresponding to the heart rate; represents terminal vascular pressure; represents distal blood pressure; Represents the speed of blood flow at the blood vessel wall.

[0037] In a second aspect of the present invention, a system for constructing a risk assessment model for myocardial infarction patients is proposed, comprising: an acquisition and access module, a data preprocessing module, an extraction and construction module, a multimodal fusion module, a personalized assessment module, an architecture management module, an adversarial detection module, an explanatory analysis module, a display and warning module, and a verification and optimization module;

[0038] The collection and access module is used to collect patient health data from various data sources such as hospital electronic medical records, wearable devices, imaging data, and laboratory test results;

[0039] The data preprocessing module is used to clean and standardize the collected data;

[0040] The multimodal fusion module is used to integrate data from different modalities into a unified feature space;

[0041] The extraction and construction module is used to extract the patient's risk characteristics and calculate hemodynamic parameters to assist in assessing the degree of vascular stenosis and the risk of thrombosis, and to screen for characteristics that affect the risk of myocardial infarction;

[0042] The personalized assessment module is used to construct a risk assessment model that adapts to individual differences among patients to predict individual myocardial infarction risks;

[0043] The architecture management module is used to optimize the data analysis of the personalized evaluation module;

[0044] The adversarial detection module is used to detect and defend against abnormal data input, preventing measurement errors or malicious attacks from affecting the prediction results;

[0045] The explanatory analysis module is used to provide explainable risk assessment results to help doctors understand the basis of the prediction results;

[0046] The display warning module is used to display risk assessment results to doctors and patients and provide corresponding health management suggestions;

[0047] The verification and optimization module is used to verify the prediction results and perform closed-loop optimization on the construction system through a continuous learning mechanism.

[0048] As a further embodiment of the present invention, the specific steps of the extraction construction module for screening features that affect the risk of myocardial infarction are as follows:

[0049] S3.1: Collect MI risk variables, including hemodynamic parameters, physiological indicators, lifestyle factors, and the probability of MI. Use the PC algorithm to automatically learn the causal structure and construct a directed acyclic causal graph based on the collected MI risk variables. Each node represents a variable that affects MI risk, and each edge represents a causal relationship.

[0050] S3.2: Consider the patient's age, sex, and history of diabetes as confounding variables. Based on each confounding variable, use Do-Calculus to calculate the probability of myocardial infarction after external intervention on each variable. Then, calculate the expected probability of myocardial infarction after each intervention and the probability of myocardial infarction without intervention. Obtain the average causal effect (ATE) of each variable based on the expected probability of myocardial infarction after intervention and the probability of myocardial infarction without intervention.

[0051] S3.3: If ATE > 0, it indicates that an increase in this variable will increase the risk of myocardial infarction. If ATE < 0, it indicates that an increase in this variable has a protective effect. Collect different treatment interventions to establish a treatment plan set and set the patient's current state as the initial state. , contains the patient's basic health data, and through the state transfer function , simulating the impact of different intervention programs on the patient's health status, Represents the current health status, represents the selected treatment plan, represents the new health status after the intervention;

[0052] S3.4: Pass ; Calculate different health status The probability of myocardial infarction under , calculate the long-term risk of the current health status, where Represents current health status The probability of myocardial infarction under represents the weight vector of health status variables, Represents the bias term, the patient's initial state As the root node, a new health state is generated according to the state transition function and used as a child node. At the same time, the treatment plan is used as an edge to connect each node to establish a corresponding solution tree, and the number of visits, state value and treatment plan reward of each node in the solution tree are initialized;

[0053] S3.5: Starting from the root node, calculate the upper confidence limit (UCB) of each child node at each level, and select the child node with the highest UCB value layer by layer. If the health state corresponding to the selected child node has not tried all treatment options, generate a new health state based on the state transition function, add the generated new health state as a child node to the tree, and evaluate its myocardial infarction probability;

[0054] S3.6: From the current node Initially, a treatment plan is randomly selected and the health status after a preset time step is simulated. The long-term risk of myocardial infarction is estimated using the state-value function. The simulation results are then backtracked to the root node, and the state values ​​and treatment plan rewards of all passed nodes are updated in sequence. The selection, expansion, simulation, and backtracking process are repeated.

[0055] S3.7: After the change in the treatment plan reward of each node converges to a preset range after multiple rounds of iteration, the iteration is stopped, and the treatment plan corresponding to the sub-node with the highest treatment plan reward is selected. A corresponding personalized treatment plan report is generated, including recommended treatment measures, expected improvement in health status, and expected reduction in myocardial infarction risk.

[0056] As a further solution of the present invention, the specific steps of the personalized assessment module to construct a risk assessment model adapted to individual differences of different patients are as follows:

[0057] S4.1: Collect data from multiple patients and create a patient dataset ,in, For the The feature vector of each patient, including static features, dynamic features and hemodynamic parameters, To represent the occurrence of myocardial infarction, a global risk assessment model was established using DNN, and its output was set as the probability of myocardial infarction. The patient data set was divided into a training set and a validation set.

[0058] S4.2: The training set is divided into multiple batches of training groups. Each training group is sequentially input into the global risk assessment model for forward propagation. The input layer receives the training group data. The hidden layer of the global risk assessment model uses multiple layers of neurons to perform nonlinear transformations on each training group. The output layer then uses a sigmoid function to calculate the probability of myocardial infarction. The cross-entropy loss function is used to calculate the loss between the model's predicted value and the true label.

[0059] S4.3: Starting from the output layer of the global risk assessment model, the loss value is back-propagated layer by layer, and the gradient of the loss value with respect to the parameters of each layer is calculated. The parameters of each layer are then updated using the Adam optimizer. The validation set is then input into the global risk assessment model, and the area under the curve (ACU) of the risk assessment model is calculated to evaluate the predictive ability of the model. If the predictive ability meets the preset standard, training is stopped. Otherwise, the global risk assessment model is repeatedly trained and verified until the model loss value converges to within the preset threshold, at which point training is stopped.

[0060] S4.4: Initialize the individual risk assessment model, inherit the parameters of the global risk assessment model, collect historical medical data of individual patients to build a small sample dataset, then input the small sample dataset into the individual risk assessment model, calculate the loss value of the individual risk assessment model, and fine-tune the parameters of the individual risk assessment model using the SGD algorithm based on the L2 regularization method;

[0061] S4.5: Test the fine-tuned individual risk assessment model using individual data not used for training to evaluate the model's personalized prediction performance. Calculate the AUC value for each individual risk assessment model to assess the model's predictive ability for individual patients. If the AUC is below the preset threshold, adjust the learning rate and retrain; otherwise, discontinue transfer learning.

[0062] S4.6: After transfer learning is complete, randomly select small sample datasets from multiple different patients and establish a task set. Then, select different patient data from the task set and use them to train the global risk assessment model. Calculate the personalized loss and, based on each personalized loss, fine-tune the individual parameters of the global risk assessment model using gradient descent.

[0063] S4.7: Calculate the sum of the losses for all fine-tuned models for all patients and optimize the model parameters. Then, test the model using new patient data that was not included in the training. Calculate the AUC value for each individual risk assessment model to assess the model's adaptability. If the AUC is below the preset threshold, adjust the learning rate and retrain. Otherwise, stop the meta-learning optimization and record the training to obtain the model parameters.

[0064] S4.8: Collect the latest patient data. If the patient has no historical medical data, calculate the similarity between the patient and the patient data stored in the medical database, and select the model parameters generated by meta-learning optimization for the corresponding patient. Use the model parameters generated by meta-learning optimization to initialize the individual risk assessment model. The global risk assessment model forward propagates the new patient data and outputs the myocardial infarction risk probability and the corresponding influencing factors through the model output layer. Conversely, directly input the new patient data into the individual risk assessment model, output the myocardial infarction risk probability and the corresponding influencing factors, and then collect doctor feedback data in real time. Through transfer learning and meta-learning optimization, the global risk assessment model and the individual risk assessment model are iteratively updated in real time.

[0065] Beneficial effects of the present invention:

[0066] This paper proposes a method and system for constructing a risk assessment model for patients with myocardial infarction. By collecting and standardizing the patient's physiological indicators, imaging data, and dynamic monitoring data, the Transformer model is used to extract time series features. These features are then mapped to a high-dimensional space through linear transformation, and self-attention weights are calculated to generate a global feature representation. Subsequently, a three-dimensional geometric model of the blood vessels is reconstructed based on coronary CTA images, the computational grid is divided, and a mathematical model of blood flow is established using the Navier-Stokes equations. Boundary conditions are set, the blood flow velocity field is solved, the wall shear stress and coronary artery stenosis rate are calculated, and coronary fluid mechanics simulation data is obtained. The high-dimensional feature matrix is ​​weightedly fused with the hemodynamic data, and a myocardial infarction risk assessment model is established based on logistic regression to calculate the probability of myocardial infarction. The SHAP method is used to visualize the contribution of each hemodynamic parameter to the prediction of myocardial infarction, determine the high-risk critical value for myocardial infarction, and provide a comprehensive risk score and high-risk factors. Subsequently, the PC algorithm was used to automatically learn the causal structure and construct a causal graph. Do-Calculus was used to calculate the impact of variable intervention on MI risk, yielding the average causal effect. If ATE > 0, it indicates an increased risk of MI; if ATE < 0, it indicates a protective effect. Treatment intervention plans were collected, a state transition function was constructed, and the impact of different intervention plans on health status was simulated. A solution tree was established to calculate the long-term MI risk for each health state. The optimal path was selected using the UCB algorithm, and repeated expansion, simulation, and backtracking were performed until convergence. Ultimately, a personalized treatment plan was generated, including recommended treatment measures, expected health improvements, and the magnitude of MI risk reduction. This optimizes the accuracy of MI risk prediction, enabling risk assessment to not only rely on statistical features but also integrate pathophysiological mechanisms, improving the clinical interpretability of predictions and avoiding the limitations of traditional data-driven methods based solely on correlation modeling. The optimal treatment plan can be dynamically adjusted, achieving data-driven personalized medicine and improving the scientific nature of intervention decisions.

[0067] The present invention proposes a method and system for constructing a risk assessment model for patients with myocardial infarction. Patient data is collected to construct a dataset containing static features, dynamic features, and hemodynamic parameters. A global risk assessment model is established using a DNN. The probability of myocardial infarction is calculated through forward propagation, the error is measured using the cross-entropy loss function, and the model parameters are updated using the Adam optimizer. Subsequently, the model's area under the curve (AUC) is evaluated on a validation set. If the model's area under the curve (AUC) does not meet the preset standard, training is continued until the loss value converges. After training is complete, an individual risk assessment model is initialized, inheriting the global model parameters, and L2 regularization and SGD fine-tuning are performed based on a small sample of individual data. The individual model's AUC is tested. If it is below a threshold, the learning rate is adjusted and retraining is performed until convergence. After transfer learning is complete, task-level training is performed based on multiple individual datasets, personalized loss is calculated, and individual parameters are optimized through gradient descent. Finally, the global model is fine-tuned. The model is tested on unseen patient data to evaluate its generalization ability. If the AUC is below a threshold, optimization is continued; otherwise, meta-learning is terminated. For new patients without historical data, the similarity between the patient and the database is calculated, and the optimized model parameters of similar patients are selected for initialization. If historical data is available, the individual risk assessment model is directly used to calculate the probability of myocardial infarction and influencing factors. This system can quickly adapt to new patient data, improve individualized prediction capabilities, increase data utilization, enhance the model's generalization ability to different patient groups, avoid training the entire model from scratch, significantly reduce computing resource consumption, compensate for insufficient individual patient data, and improve prediction capabilities for rare cases. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The present invention will be further described below with reference to the accompanying drawings.

[0069] Figure 1 A flowchart of a method for constructing a risk assessment model for myocardial infarction patients provided in an embodiment of the present invention;

[0070] Figure 2 This is a framework diagram of a system for constructing a risk assessment model for myocardial infarction patients provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0071] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0072] Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative work shall fall within the scope of protection of the present invention.

[0073] The present invention provides a method and system for constructing a risk assessment model for myocardial infarction patients. Figure 1 , Figure 1 A flowchart of a method for constructing a risk assessment model for myocardial infarction patients provided in an embodiment of the present invention. The method comprises the following steps:

[0074] Collect data from various modalities of hospital information systems, image archiving systems, and wearable devices, and integrate data from different modalities.

[0075] Specifically, the collected electrocardiogram (ECG), coronary artery CTA, and laboratory examination data are standardized and normalized by Z-score, and the data of each modality are shrunk to the preset data interval. The time series information in each modality data is aligned by linear interpolation, and the sampling frequency of each modality data is synchronized by the sliding window method. The data of each modality are traversed, and the missing values ​​in each modality data are processed by mean filling. Then, each data is tested to see whether it deviates from 3 times the standard deviation of the mean. If it deviates, the corresponding data is judged as an outlier and is eliminated or replaced by the mean. The noise information in each modality data is then removed by wavelet transform. CNN and LSTM encoding networks are used to extract the special information of different modal data and map them to the same feature space to obtain the single modal features of each modal data in the same feature space. Different weights are assigned to each single modal feature based on the contribution of different modalities. Then, all single modal features are weighted and combined to form a global feature vector.

[0076] In this embodiment, the specific calculation formula for Z-score standardization is as follows:

[0077] ;

[0078] ;

[0079] Where, represents the mean of the data; represents the total number of data; Representative Original data values; represents the standard deviation of the data; Representative The standardized data values;

[0080] The specific calculation formula for the normalization process described in S1.1 is as follows: ;

[0081] Where, Representative Normalized data values; Representative Original data values; Represents the maximum value in multimodal data; Represents the minimum value in multimodal data.

[0082] Extract patient characteristic data, analyze the impact of hemodynamic parameters on the occurrence of myocardial infarction, and identify key factors affecting the risk of myocardial infarction.

[0083] Specifically, the generated global feature vectors are collected to construct a multimodal feature set ,in Indicates the The feature vector of time steps, and , by using linear transformation to map each input data in the multimodal feature to a high-dimensional space to obtain an input embedding vector, wherein the query, key and value matrices of each input embedding vector are calculated, and based on the Query matrix and the Key matrix, the self-attention weights between the input embedding vectors at different time steps are calculated through different attention heads, the attention scores are normalized by the Softmax function, and the weighted sum of the attention of each input embedding vector is calculated, and then the weighted sum results of each attention head are concatenated to generate the output features, and the output features of each time step are nonlinearly transformed by the FFN network, and the output features after nonlinear transformation are layer-normalized, and then the original global feature vector is added to the output features to generate the corresponding high-dimensional feature matrix, and then the vascular structure information is obtained through the coronary artery CTA imaging data, and the vascular centerline is set as , the vessel radius is set to ,in The arc length coordinates along the center line of the blood vessel are used. According to the obtained vascular structure information, a three-dimensional geometric model of the coronary artery is constructed, and the computational grid is divided. The blood flow mathematical model in the coronary artery is established through the Navier-Stokes equation. The coronary artery inlet boundary conditions, outlet boundary conditions and vascular wall boundary conditions are set according to the established blood flow mathematical model. By solving the blood flow mathematical model, the blood flow velocity field at different positions in the blood vessel is obtained, and the maximum flow velocity in the stenosis area of ​​the blood vessel is calculated. According to the blood flow velocity field at different positions and the three-dimensional geometric model of the coronary artery, the wall shear stress and the coronary artery stenosis rate are calculated respectively to obtain the fluid mechanics simulation data of each coronary artery. The generated high-dimensional feature matrix coronary dynamics The vascular fluid dynamics simulation data were weighted and fused to form a global feature representation. A myocardial infarction risk assessment model was established through logistic regression. The probability of myocardial infarction was calculated using this myocardial infarction risk assessment model. Then, based on the probability of myocardial infarction of each hemodynamic parameter, the characteristic contribution of different hemodynamic parameters in the model was calculated, and the impact of the numerical range of hemodynamic parameters on the occurrence of myocardial infarction was analyzed. The critical value of high risk of myocardial infarction was calculated. The probability of myocardial infarction, the characteristic contribution of different hemodynamic parameters in the model, and the critical value of high risk of myocardial infarction were visualized using the SHAP method to demonstrate the impact of each hemodynamic parameter on the prediction of myocardial infarction, give the patient a comprehensive risk score, and mark high-risk factors.

[0084] Specifically, the linear change calculation formula is as follows: Where, Represents the time step Input feature embedding of Represents a trainable weight matrix; Represents the time step The input data, ; represents the bias term;

[0085] The specific calculation formula for the weighted sum of input embedding vector attention is as follows:

[0086] ;

[0087] ;

[0088] ;

[0089] ;

[0090] ;

[0091] Where, Represents input feature embedding; represents the query matrix; represents the bond matrix; Representative value matrix; 、 as well as Represent a set of trainable weight matrices respectively; Represents the attention score matrix, which is used to compare the correlation between features at different time steps; ; represents the dimension of the bond matrix; Represents the feature matrix after self-attention calculation; represents the normalization function; Represents multi-head attention output; Representative The result of an attention head; represents the sum function; represents the linear transformation parameter;

[0092] The specific expression of the blood flow mathematical model is as follows: Where, represents blood density; represents the blood flow velocity vector; Represents the current time; represents the velocity gradient; Represents blood pressure; represents blood dynamic viscosity; Laplace operator representing velocity; since coronary blood flow is pulsating, the momentum equation needs to be combined with the continuity equation , to ensure; the specific boundary conditions of the blood flow mathematical model are as follows: ;

[0093] in, Representative Moment coronary artery inlet flow velocity; represents the average flow velocity; represents the pulsation amplitude; Represents the basic frequency corresponding to the heart rate; represents terminal vascular pressure; represents distal blood pressure; Represents the speed of blood flow at the blood vessel wall.

[0094] Train personalized risk assessment models, build lightweight edge risk prediction models based on real-time data streams, and perform personalized risk predictions for different patients.

[0095] Detect potentially abnormal samples, compare risk assessment results across different patient groups, analyze the model's adaptability in different scenarios, and continuously optimize model parameters based on clinical expert feedback.

[0096] Based on the same inventive concept, the present invention also provides a system for constructing a risk assessment model for myocardial infarction patients. Figure 2 , Figure 2 A schematic diagram of the structure of a system for constructing a risk assessment model for myocardial infarction patients provided in an embodiment of the present invention, comprising: an acquisition and access module, a data preprocessing module, an extraction and construction module, a multimodal fusion module, a personalized assessment module, an architecture management module, an adversarial detection module, an explanatory analysis module, a display and warning module, and a verification and optimization module;

[0097] The acquisition and access module is used to collect patients' health data from various data sources such as hospital electronic medical records, wearable devices, imaging data, and laboratory test results; the data preprocessing module is used to clean and standardize the collected data; the multimodal fusion module is used to integrate data from different modalities into a unified feature space; the extraction and construction module is used to extract patients' risk characteristics and calculate hemodynamic parameters to assist in assessing the degree of vascular stenosis and the risk of thrombosis, and to screen features that affect the risk of myocardial infarction.

[0098] Specifically, various myocardial infarction risk variables such as hemodynamic parameters, physiological indicators, lifestyle factors, and the probability of myocardial infarction are collected, and the causal structure is automatically learned through the PC algorithm. A directed acyclic causal graph is constructed based on the collected myocardial infarction risk variables, in which each node represents a variable that affects the risk of myocardial infarction, and the edge represents the causal relationship. The patient's age, gender, and diabetes history are used as confounding variables. Based on each confounding variable, Do-Calculus is used to calculate the probability of myocardial infarction after external intervention on different variables. The expected probability of myocardial infarction after each variable intervention and the probability of myocardial infarction without intervention are then calculated. Based on the expected probability of myocardial infarction after intervention and the probability of myocardial infarction without intervention, the average causal effect ATE of each variable is obtained. If ATE>0, it indicates that an increase in the variable will increase the risk of myocardial infarction. If ATE<0, it indicates that an increase in the variable has a protective effect. Different treatment interventions are collected to establish a treatment plan set, and the patient's current state is set as the initial state. , contains the patient's basic health data, and through the state transfer function , simulating the impact of different intervention programs on the patient's health status, Represents the current health status, represents the selected treatment plan, represents the new health status after the intervention, ; Calculate different health status The probability of myocardial infarction under , calculate the long-term risk of the current health status, where Represents current health status The probability of myocardial infarction under represents the weight vector of health status variables, Represents the bias term, the patient's initial state As the root node, a new health state is generated according to the state transfer function, and it is used as a child node. At the same time, the treatment plan is used as an edge to connect each node to establish a corresponding solution tree, and the number of visits, state value and treatment plan reward of each node in the solution tree are initialized. Starting from the root node, the confidence upper limit value UCB of each layer of child nodes is calculated, and the child node with the highest UCB value is selected layer by layer. If the health state corresponding to the selected child node has not tried all treatment plans, a new health state is generated according to the state transfer function, and the generated new health state is added to the tree as a child node, and its myocardial infarction probability is evaluated, starting from the current node Initially, a treatment plan is randomly selected, and the health status after a preset time step is simulated. The long-term risk of myocardial infarction is estimated through the state value function, and the simulation results are traced back to the root node. The state value and treatment plan reward of all passed nodes are updated in turn. The selection, expansion, simulation and backtracking process are repeated until the change value of the treatment plan reward of each node converges to the preset range after multiple rounds of iterations. The iteration is stopped, and the treatment plan corresponding to the child node with the highest treatment plan reward is selected to generate a corresponding personalized treatment plan report, including recommended treatment measures, expected improvement in health status and expected reduction in myocardial infarction risk.

[0099] The personalized assessment module is used to build a risk assessment model that adapts to the individual differences of different patients to predict individual myocardial infarction risk.

[0100] Specifically, collect data from multiple patients and establish a patient dataset ,in, For the The feature vector of each patient, including static features, dynamic features and hemodynamic parameters, Representing the occurrence of myocardial infarction, a global risk assessment model is established using DNN, and its output is set as the probability of myocardial infarction. The patient data set is divided into a training set and a validation set, and the training set is divided into multiple batches of training groups. Each training group is input into the global risk assessment model in turn for forward propagation. The input layer receives the training group data. The hidden layer in the global risk assessment model uses multi-layer neurons to perform nonlinear transformation on each training group. After that, the output layer uses the Sigmoid function to calculate the probability of myocardial infarction, and the cross-entropy loss function is used to calculate the loss value of the model prediction value and the true label. The loss value starts from the output layer of the global risk assessment model, and is back-propagated layer by layer. The gradient of the loss value for the parameters of each layer is calculated, and then the Adam optimizer is used to optimize each layer. The layer parameters are updated, and then the validation set is input into the global risk assessment model, and the area under the curve ACU of the risk assessment model is calculated to evaluate the predictive ability of the model. If the predictive ability reaches the preset standard, the training is stopped. Otherwise, the global risk assessment model is repeatedly trained and verified until the model loss value converges to the preset threshold. The training is stopped, the individual risk assessment model is initialized, and the parameters of the global risk assessment model are inherited. The historical medical data of individual patients are collected to establish a small sample data set. The small sample data set is then input into the individual risk assessment model, and the loss value of the individual risk assessment model is calculated. Based on the L2 regularization method, the parameters of the individual risk assessment model are fine-tuned by the SGD algorithm, and the individual data not used for training are used to fine-tune the micro risk assessment model. The adjusted individual risk assessment model is tested to evaluate the personalized prediction performance of the model, and the ACU value of each individual risk assessment model is calculated to evaluate the predictive ability of the model on individual patients. If the AUC is lower than the preset threshold, the learning rate is adjusted and retraining is performed. Otherwise, the transfer learning is stopped. After the transfer learning is completed, a small sample data set of multiple different patients is randomly selected, and a task set is established. Then, different patient data are selected from the task set, and the global risk assessment model is trained using different patient data, and the personalized loss is calculated. Based on each personalized loss, the individual parameters of the global risk assessment model are fine-tuned by gradient descent, the sum of the losses of the fine-tuned models of all patients is calculated, and the model parameters are optimized. Then, new patients who did not participate in the training are used. The patient data is tested and the ACU value of each individual risk assessment model is calculated to evaluate the adaptability of the model. If the AUC is lower than the preset threshold, the learning rate is adjusted and retraining is performed. Otherwise, the meta-learning optimization is stopped, and the training is recorded to obtain the parameters of each model. The latest patient data is collected. If the patient has no historical medical data, the similarity between the patient and the patient data stored in the medical database is calculated, and the model parameters generated by the meta-learning optimization of the corresponding patient are selected. The model parameters generated by the meta-learning optimization are initialized with the individual risk assessment model. The global risk assessment model performs forward propagation on the new patient data and outputs the myocardial infarction risk probability and the corresponding influencing factors through the model output layer. Otherwise, the new patient data is directly input into the individual risk assessment model.Output the risk probability of myocardial infarction and the corresponding influencing factors, then collect doctor feedback data in real time, and iterate and update the global risk assessment model and individual risk assessment model in real time through transfer learning and meta-learning optimization.

[0101] The architecture management module is used to optimize the data analysis of the personalized assessment module; the adversarial detection module is used to detect and defend against abnormal data input to prevent measurement errors or malicious attacks from affecting the prediction results; the explanatory analysis module is used to provide explainable risk assessment results to help doctors understand the basis of the prediction results; the display and warning module is used to display the risk assessment results to doctors and patients and provide corresponding health management recommendations; the verification and optimization module is used to verify the prediction results and perform closed-loop optimization of the construction system through a continuous learning mechanism.

[0102] The above is a detailed description of an embodiment of the present invention. However, the content described is only a preferred embodiment of the present invention and should not be considered to limit the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A method for constructing a risk assessment model for myocardial infarction patients, characterized in that: The following steps are involved: Ⅰ. Collect data from hospital information systems, imaging archiving systems, and wearable devices, and integrate data from different modalities; II. Extract patient characteristic data, analyze the impact of hemodynamic parameters on the occurrence of myocardial infarction, and identify key factors affecting the risk of myocardial infarction; III. Train a personalized risk assessment model and build a lightweight edge risk prediction model based on real-time data streams to perform personalized risk prediction for different patients; IV. Detect potentially abnormal samples, compare risk assessment results across different patient groups, analyze the model's adaptability in different scenarios, and continuously optimize model parameters based on clinical expert feedback; The specific steps for integrating data from different modalities described in Step I are as follows: S1.1: Z-score and normalize the collected ECG, coronary artery CT scan, and laboratory test data to shrink each modality to a preset data interval. Linear interpolation is used to align the time series information in each modality, and the sliding window method is used to synchronize the sampling frequency of each modality. S1.2: Traverse each modal data and process the missing values ​​in each modal data by filling the mean. Then check whether each data deviates from the mean by three times the standard deviation. If it deviates, the corresponding data is judged as an outlier and is removed or replaced with the mean. Then, the wavelet transform is used to remove the noise information in each modal data. S1.3: Use CNN and LSTM encoding networks to extract the unique information of different modal data and map them to the same feature space to obtain the unimodal features of each modal data in the same feature space. Based on the contribution of different modalities, different weights are assigned to each unimodal feature. All unimodal features are then weighted and combined to form a global feature vector. The specific steps for analyzing the effects of hemodynamic parameters on the occurrence of myocardial infarction in step II are as follows: S2.1: Collect the generated global feature vectors to construct a multimodal feature set ,in Indicates the The feature vector of time steps, and , by using linear transformation to map each input data in the multimodal feature to a high-dimensional space to obtain the input embedding vector, where; S2.2: Calculate the query, key, and value matrices for each input embedding vector. Based on the query and key matrices, calculate the self-attention weights between the input embedding vectors at different time steps through different attention heads. Normalize the attention scores using the Softmax function, and calculate the weighted sum of the attention of each input embedding vector. Then, concatenate the weighted sum results of each attention head to generate the output features. S2.3: Perform nonlinear transformation on the output features of each time step through the FFN network, perform layer normalization on the output features after nonlinear transformation, add the original global feature vector to the output features to generate the corresponding high-dimensional feature matrix, and then obtain the vascular structure information through the coronary artery CTA image data, and set the vascular centerline as , the vessel radius is set to ,in is the arc length coordinate along the centerline of the blood vessel; S2.4: Construct a three-dimensional geometric model of the coronary arteries based on the acquired vascular structure information and divide the computational grid. Establish a mathematical model of blood flow in the coronary arteries using the Navier-Stokes equations. Set boundary conditions for the coronary artery inlet and outlet, as well as the vessel wall, based on the established blood flow mathematical model. S2.5: By solving the blood flow mathematical model, the blood flow velocity field at different locations within the blood vessel is obtained, and the maximum flow velocity in the stenotic region of the vessel is calculated. Based on the blood flow velocity field at different locations and the three-dimensional geometric model of the coronary artery, the wall shear stress and coronary artery stenosis rate are calculated to obtain fluid dynamics simulation data for each coronary artery; S2.6: Perform weighted fusion on the generated high-dimensional feature matrix and coronary artery fluid dynamics simulation data to form a global feature representation. Establish a myocardial infarction risk assessment model through logistic regression. Use this model to calculate the probability of myocardial infarction. Then, based on the probability of myocardial infarction for each hemodynamic parameter, calculate the characteristic contribution of different hemodynamic parameters in the model. Analyze the impact of the numerical range of hemodynamic parameters on the occurrence of myocardial infarction, and calculate the critical value for high risk of myocardial infarction. S2.7: Use the SHAP method to visualize the probability of myocardial infarction, the characteristic contribution of different hemodynamic parameters in the model, and the critical value of high risk of myocardial infarction, demonstrate the influence of each hemodynamic parameter on the prediction of myocardial infarction, give the patient's comprehensive risk score, and mark high-risk factors.

2. The method for constructing a risk assessment model for myocardial infarction patients according to claim 1, characterized in that: The specific calculation formula for Z-score standardization described in S1.1 is as follows: ； ； ； Where, represents the mean of the data; represents the total number of data; Representative Original data values; represents the standard deviation of the data; Representative The standardized data values; The specific calculation formula for the normalization process described in S1.1 is as follows: ; Where, Representative Normalized data values; Representative Original data values; Represents the maximum value in multimodal data; Represents the minimum value in multimodal data.

3. The method for constructing a risk assessment model for myocardial infarction patients according to claim 1, characterized in that: The specific calculation formula for the linear change described in S2.1 is as follows: ; Where, Represents the time step Input feature embedding of Represents a trainable weight matrix; Represents the time step The input data, ; represents the bias term; The specific calculation formula for the weighted sum of the input embedding vector attention described in S2.2 is as follows: ； ； ； ； ； ； Where, Represents input feature embedding; represents the query matrix; represents the bond matrix; Representative value matrix; 、 as well as Represent a set of trainable weight matrices respectively; Represents the attention score matrix, which is used to compare the correlation between features at different time steps; Representative The attention head The key matrix at the moment; represents the dimension of the bond matrix; Represents the feature matrix after self-attention calculation; represents the normalization function; Represents multi-head attention output; Representative The result of an attention head; represents the sum function; represents the linear transformation parameter; The specific expression of the blood flow mathematical model described in S2.4 is as follows: ； Where, represents blood density; represents the blood flow velocity vector; Represents the current time; represents the velocity gradient; Represents blood pressure; represents blood dynamic viscosity; Laplace operator representing velocity; since coronary blood flow is pulsating, the momentum equation needs to be combined with the continuity equation , to ensure; the specific boundary conditions of the blood flow mathematical model are as follows: ; in, Representative Moment coronary artery inlet flow velocity; represents the average flow velocity; Representative pulse Dynamic amplitude; Represents the basic frequency corresponding to the heart rate; represents terminal vascular pressure; represents distal blood pressure; Represents the speed of blood flow at the blood vessel wall.

4. A system for constructing a risk assessment model for myocardial infarction patients, for implementing the method for constructing a risk assessment model for myocardial infarction patients according to any one of claims 1 to 3, characterized in that: include: Collection and access module, data preprocessing module, extraction and construction module, multimodal fusion module, personalized evaluation module, architecture management module, adversarial detection module, explanatory analysis module, display and warning module, and verification and optimization module; The collection and access module is used to collect patient health data from various data sources such as hospital electronic medical records, wearable devices, imaging data, and laboratory test results; The data preprocessing module is used to clean and standardize the collected data; The multimodal fusion module is used to integrate data from different modalities into a unified feature space; The extraction and construction module is used to extract the patient's risk characteristics and calculate hemodynamic parameters to assist in assessing the degree of vascular stenosis and the risk of thrombosis, and to screen for characteristics that affect the risk of myocardial infarction; The personalized assessment module is used to construct a risk assessment model that adapts to individual differences among patients to predict individual myocardial infarction risks; The architecture management module is used to optimize the data analysis of the personalized evaluation module; The adversarial detection module is used to detect and defend against abnormal data input, preventing measurement errors or malicious attacks from affecting the prediction results; The explanatory analysis module is used to provide explainable risk assessment results to help doctors understand the basis of the prediction results; The display warning module is used to display risk assessment results to doctors and patients and provide corresponding health management suggestions; The verification and optimization module is used to verify the prediction results and perform closed-loop optimization on the construction system through a continuous learning mechanism.

5. The system for constructing a risk assessment model for myocardial infarction patients according to claim 4, characterized in that: The specific steps of the extraction construction module to screen features that affect the risk of myocardial infarction are as follows: S3.1: Collect MI risk variables, including hemodynamic parameters, physiological indicators, lifestyle factors, and the probability of MI. Use the PC algorithm to automatically learn the causal structure and construct a directed acyclic causal graph based on the collected MI risk variables. Each node represents a variable that affects MI risk, and each edge represents a causal relationship. S3.2: Consider the patient's age, sex, and history of diabetes as confounding variables. Based on each confounding variable, use Do-Calculus to calculate the probability of myocardial infarction after external intervention on each variable. Then, calculate the expected probability of myocardial infarction after each intervention and the probability of myocardial infarction without intervention. Obtain the average causal effect (ATE) of each variable based on the expected probability of myocardial infarction after intervention and the probability of myocardial infarction without intervention. S3.3: If ATE > 0, it indicates that an increase in this variable will increase the risk of myocardial infarction. If ATE < 0, it indicates that an increase in this variable has a protective effect. Collect different treatment interventions to establish a treatment plan set and set the patient's current state as the initial state. , contains the patient's basic health data, and through the state transfer function , simulating the impact of different intervention programs on the patient's health status, Represents the current health status, represents the selected treatment plan, represents the new health status after the intervention; S3.4: Pass Calculate different health status The probability of myocardial infarction under , calculate the long-term risk of the current health status, where Represents current health status The probability of myocardial infarction under represents the weight vector of health status variables, Represents the bias term, the patient's initial state As the root node, a new health state is generated according to the state transition function and used as a child node. At the same time, the treatment plan is used as an edge to connect each node to establish a corresponding solution tree, and the number of visits, state value and treatment plan reward of each node in the solution tree are initialized; S3.5: Starting from the root node, calculate the upper confidence limit (UCB) of each child node at each level, and select the child node with the highest UCB value layer by layer. If the health state corresponding to the selected child node has not tried all treatment options, generate a new health state based on the state transition function, add the generated new health state as a child node to the tree, and evaluate its myocardial infarction probability; S3.6: From the current node Initially, a treatment plan is randomly selected and the health status after a preset time step is simulated. The long-term risk of myocardial infarction is estimated using the state-value function. The simulation results are then backtracked to the root node, and the state values ​​and treatment plan rewards of all passed nodes are updated in sequence. The selection, expansion, simulation, and backtracking process are repeated. S3.7: After the change in the treatment plan reward of each node converges to a preset range after multiple rounds of iteration, the iteration is stopped, and the treatment plan corresponding to the sub-node with the highest treatment plan reward is selected. A corresponding personalized treatment plan report is generated, including recommended treatment measures, expected improvement in health status, and expected reduction in myocardial infarction risk.

6. The system for constructing a risk assessment model for myocardial infarction patients according to claim 4, characterized in that: The specific steps of constructing a risk assessment model adapted to individual differences of different patients by the personalized assessment module are as follows: S4.1: Collect data from multiple patients and create a patient dataset ,in, For the The feature vector of each patient, including static features, dynamic features and hemodynamic parameters, To represent the occurrence of myocardial infarction, a global risk assessment model was established using DNN, and its output was set as the probability of myocardial infarction. The patient data set was divided into a training set and a validation set. S4.2: The training set is divided into multiple batches of training groups. Each training group is sequentially input into the global risk assessment model for forward propagation. The input layer receives the training group data. The hidden layer of the global risk assessment model uses multiple layers of neurons to perform nonlinear transformations on each training group. The output layer then uses a sigmoid function to calculate the probability of myocardial infarction. The cross-entropy loss function is used to calculate the loss between the model's predicted value and the true label. S4.3: Starting from the output layer of the global risk assessment model, the loss value is back-propagated layer by layer, and the gradient of the loss value with respect to the parameters of each layer is calculated. The parameters of each layer are then updated using the Adam optimizer. The validation set is then input into the global risk assessment model, and the area under the curve (ACU) of the risk assessment model is calculated to evaluate the predictive ability of the model. If the predictive ability meets the preset standard, training is stopped. Otherwise, the global risk assessment model is repeatedly trained and verified until the model loss value converges to within the preset threshold, at which point training is stopped. S4.4: Initialize the individual risk assessment model, inherit the parameters of the global risk assessment model, collect historical medical data of individual patients to build a small sample dataset, then input the small sample dataset into the individual risk assessment model, calculate the loss value of the individual risk assessment model, and fine-tune the parameters of the individual risk assessment model using the SGD algorithm based on the L2 regularization method; S4.5: Test the fine-tuned individual risk assessment model using individual data not used for training to evaluate the model's personalized prediction performance. Calculate the AUC value for each individual risk assessment model to assess the model's predictive ability for individual patients. If the AUC is below the preset threshold, adjust the learning rate and retrain; otherwise, discontinue transfer learning. S4.6: After transfer learning is complete, randomly select small sample datasets from multiple different patients and establish a task set. Then, select different patient data from the task set and use them to train the global risk assessment model. Calculate the personalized loss and, based on each personalized loss, fine-tune the individual parameters of the global risk assessment model using gradient descent. S4.7: Calculate the sum of the losses for all fine-tuned models for all patients and optimize the model parameters. Then, test the model using new patient data that was not included in the training. Calculate the AUC value for each individual risk assessment model to assess the model's adaptability. If the AUC is below the preset threshold, adjust the learning rate and retrain. Otherwise, stop the meta-learning optimization and record the training to obtain the model parameters. S4.8: Collect the latest patient data. If the patient has no historical medical data, calculate the similarity between the patient and the patient data stored in the medical database, and select the model parameters generated by meta-learning optimization for the corresponding patient. Use the model parameters generated by meta-learning optimization to initialize the individual risk assessment model. The global risk assessment model forward propagates the new patient data and outputs the myocardial infarction risk probability and the corresponding influencing factors through the model output layer. Conversely, directly input the new patient data into the individual risk assessment model, output the myocardial infarction risk probability and the corresponding influencing factors, and then collect doctor feedback data in real time. Through transfer learning and meta-learning optimization, the global risk assessment model and the individual risk assessment model are iteratively updated in real time.

Images