Artificial intelligence-based coronary heart disease prediction model generation method

Abstract

The application relates to the field of artificial intelligence, and discloses a coronary heart disease prediction model generation method based on artificial intelligence, which comprises the following steps: acquiring dynamic physiological time series data and static blood vessel structure data of a test object; respectively extracting multi-scale dynamic physiological representation and topological blood vessel structure representation; performing cross-modal space-time alignment and feature interaction through a bidirectional cross-attention mechanism to generate fusion representation; and training a coronary heart disease risk prediction model based on the fusion representation. The method comprises the following modules: multi-source data acquisition, dynamic physiological feature coding, blood vessel structure graph modeling, cross-modal feature fusion, and prediction model training. The application synchronously integrates dynamic physiological time series data continuously collected for a long time and high-resolution coronary artery structure image data, and constructs a coronary heart disease risk assessment system in dual dimensions of physiological function and anatomical structure.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a method for generating a coronary heart disease prediction model based on artificial intelligence. Background Technology

[0002] With the deepening application of artificial intelligence technology in the healthcare field, data-driven disease prediction models have become a key means to improve the early screening and intervention capabilities for cardiovascular diseases. Coronary artery disease (CAD), as one of the world's leading causes of death from chronic diseases, exhibits insidious and progressive pathological development. Clinical diagnosis has long relied on static imaging examinations such as coronary angiography, electrocardiograms, or coronary CTA. While these methods can reflect anatomical abnormalities at specific time points, they struggle to capture the evolution of dynamic physiological parameters during disease progression and their complex relationship with vascular network topology, thus limiting the ability to identify lesions in the compensated or subclinical stages.

[0003] Intelligent prediction methods based on the fusion of multi-source physiological signals and vascular structural information have become a research hotspot. This direction aims to integrate continuously monitored temporal physiological data (such as heart rate variability, blood pressure fluctuations, and blood oxygen saturation) with the anatomical connectivity of coronary arteries. By modeling the function-structure coupling mechanism between organs and vascular segments, it reveals systemic compensatory signals preceding the onset of coronary artery disease. The core of such methods lies in how to effectively characterize the propagation and interaction patterns of dynamic physiological states on the vascular network and extract spatiotemporal features with predictive value.

[0004] Processing time-series signals and structural data separately, or using simple concatenation or weighted averaging for later fusion, fails to establish an explicit correlation between physiological dynamics and vascular topology. On one hand, traditional machine learning models lack the ability to model vascular network structures in non-Euclidean space, failing to characterize hemodynamics or the conduction characteristics of electrical signals along vascular pathways. On the other hand, while mainstream deep learning architectures can process time-series data, they neglect the differences in the contribution of different time points to disease prediction and lack a dynamic focusing mechanism for key compensatory windows. Especially in real clinical scenarios, with significant individual patient differences, high noise in monitoring data, and strong heterogeneity in vascular structures, these deficiencies lead to weak model generalization ability and low early warning sensitivity, making it difficult to meet the urgent need for highly robust and interpretable predictive tools in precision medicine. Summary of the Invention

[0005] This invention provides a method for generating coronary artery disease (CAD) prediction models based on artificial intelligence, aiming to solve the technical problems of traditional CAD prediction models that rely on single static data and ignore the deep correlation between dynamic physiological temporal relationships and vascular structural features. This method constructs a multimodal fusion architecture to spatiotemporally align and feature-couple continuous dynamic physiological signals with static vascular anatomy, thereby generating a prediction model that can accurately capture the early pathological evolution of CAD.

[0006] This invention provides a method for generating a coronary heart disease prediction model based on artificial intelligence, comprising: acquiring dynamic physiological time-series data and static vascular structure data of the subject; extracting multi-scale temporal features from the dynamic physiological time-series data to generate a dynamic physiological representation; performing topological graph modeling on the static vascular structure data to generate a vascular structure representation; performing cross-modal spatiotemporal alignment and feature interaction between the dynamic physiological representation and the vascular structure representation to generate a fused representation; training a coronary heart disease risk prediction model based on the fused representation, and outputting the probability of coronary heart disease occurrence.

[0007] As one embodiment of the present invention, the acquisition of dynamic physiological time-series data and static vascular structure data of the subject specifically includes: continuously collecting electrocardiogram signals for no less than 24 hours using an electrocardiogram monitoring device, extracting heart rate variability sequences from them, wherein the heart rate variability sequences include the standard deviation of adjacent normal heartbeat intervals, the root mean square of the difference between adjacent heartbeat intervals, and the ratio of low-frequency power to high-frequency power; simultaneously collecting systolic and diastolic blood pressure time series for no less than 24 hours using a non-invasive blood pressure monitoring device, calculating blood pressure variability indicators, wherein the blood pressure variability indicators include the standard deviation of systolic blood pressure, the coefficient of variation of diastolic blood pressure, and the amplitude of pulse pressure fluctuation; acquiring three-dimensional image data of the coronary arteries using a coronary computed tomography angiography device, reconstructing the coronary artery tree structure, and marking the lumen diameter, degree of stenosis, plaque distribution location, and calcification score of each major branch.

[0008] As one embodiment of the present invention, the multi-scale temporal feature extraction of dynamic physiological time-series data to generate dynamic physiological representation specifically includes: dividing the heart rate variability sequence and blood pressure time series into multiple subsequences according to a preset time window, with each time window having a length of 5 minutes and a step size of 1 minute; inputting each subsequence into a one-dimensional convolutional neural network, which contains three convolutional layers with kernel sizes of 3, 5, and 7, and channel numbers of 64, 128, and 256, respectively, followed by batch normalization and modified linear unit activation functions after each layer; performing temporal modeling of the convolutional output of each time window through a bidirectional gated recurrent unit with a hidden unit number of 512; and finally outputting a dynamic physiological representation vector with a dimension of 1024, which encodes the nonlinear coupled dynamic features of heart rate and blood pressure at short-term, medium-term, and long-term time scales.

[0009] As one embodiment of the present invention, the method of generating a vascular structure representation by performing topological graph modeling on static vascular structure data specifically includes: processing coronary artery three-dimensional image data using a centerline extraction algorithm to generate a directed graph composed of nodes and edges, where nodes represent vascular bifurcation points or endpoints, and edges represent vascular segments; assigning an attribute vector to each node, the attribute vector containing the average lumen diameter, local curvature, and estimated wall shear stress at that node; assigning an attribute vector to each edge, the attribute vector containing the length of that vascular segment, average stenosis rate, plaque load percentage, and calcification density; inputting the directed graph into a graph neural network, which employs a three-layer graph attention mechanism, calculating the attention weights between nodes when aggregating neighborhood information in each layer, with 8 attention heads and an output dimension of 256; and generating a vascular structure representation vector with a dimension of 256 after global max pooling operation.

[0010] As one embodiment of the present invention, the cross-modal spatiotemporal alignment and feature interaction of dynamic physiological representation and vascular structure representation specifically includes: mapping the dynamic physiological representation vector to the same embedding space as the vascular structure representation through a fully connected layer to obtain the aligned dynamic representation; constructing a cross-attention module, using the vascular structure representation as the query vector and the aligned dynamic representation as the key-value vector, to calculate the structure-guided dynamic feature response; simultaneously constructing another cross-attention module, using the dynamic physiological representation as the query vector and the vascular structure representation as the key-value vector, to calculate the dynamically-guided structural feature response; concatenating the outputs of the two cross-attention modules, and generating a fused representation with a dimension of 1536 through residual connections and layer normalization.

[0011] As one embodiment of the present invention, the training of the coronary heart disease risk prediction model based on fusion representation specifically includes: inputting the fusion representation into a multilayer perceptron classifier, which contains three fully connected layers with hidden units of 1000, 500, and 100 respectively, and the activation function being a Gaussian error linear unit; the last layer outputs a single scalar, which is mapped to the probability of coronary heart disease occurrence via a Sigmoid function; the model training uses a focus loss function, focusing on hard-to-classify samples, and the adjustment factor in the loss function... The value is 2, which is the balance parameter. The value is set to 0.75; 5-fold cross-validation is used during training, the optimizer is AdamW, the initial learning rate is 0.001, and the weight decay coefficient is 0.01.

[0012] According to another aspect of the present invention, an artificial intelligence-based coronary heart disease prediction model generation system is provided, comprising: a multi-source data acquisition module for acquiring dynamic physiological time-series data and static vascular structure data of the subject; a dynamic physiological feature encoding module for extracting multi-scale temporal features from the dynamic physiological time-series data to generate dynamic physiological representations; a vascular structure graph modeling module for performing topological graph modeling on the static vascular structure data to generate vascular structure representations; a cross-modal feature fusion module for performing cross-modal spatiotemporal alignment and feature interaction between the dynamic physiological representations and the vascular structure representations to generate fused representations; and a prediction model training module for training a coronary heart disease risk prediction model based on the fused representations and outputting the probability of coronary heart disease occurrence.

[0013] In one embodiment of the present invention, the multi-source data acquisition module includes an electrocardiogram (ECG) signal acquisition unit, a blood pressure signal acquisition unit, and a coronary artery imaging acquisition unit; the ECG signal acquisition unit is configured to continuously record standard 12-lead ECG signals with a sampling frequency of not less than 250 Hz; the blood pressure signal acquisition unit is configured to automatically measure blood pressure every 15 minutes using oscillometric methods and simultaneously record continuous pulse waveforms with a sampling frequency of not less than 100 Hz; the coronary artery imaging acquisition unit is configured to use a multi-slice spiral computed tomography (CT) scanner with a slice thickness not exceeding 0.6 mm and a reconstruction matrix of 512×512.

[0014] As one embodiment of the present invention, the dynamic physiological feature encoding module integrates a one-dimensional convolutional neural network and a bidirectional gated recurrent unit. The one-dimensional convolutional neural network has 5 input channels, corresponding to three heart rate variability indicators and two indicators: systolic blood pressure and diastolic blood pressure. The forward and backward hidden states of the bidirectional gated recurrent unit are concatenated in the time dimension and compressed into a fixed-length vector through time pooling.

[0015] As one embodiment of the present invention, the vascular structure modeling module integrates a centerline extraction algorithm unit and a graph neural network unit; the centerline extraction algorithm unit uses a fast traversal method combined with minimum path search to extract the centerlines of the main trunk and branches from the three-dimensional image of the coronary artery; the graph neural network unit is configured to receive attribute tensors of nodes and edges and execute a message passing mechanism, fusing its own features and the weighted features of neighboring nodes when updating node embedding at each layer.

[0016] In one embodiment of the present invention, the cross-modal feature fusion module integrates two cross-attention sub-modules and a feature splicing unit; each cross-attention sub-module includes a query projection matrix, a key projection matrix and a value projection matrix, all with a projection dimension of 256; the feature splicing unit splices the outputs of the two sub-modules along the feature dimension and connects them to the residual connection path, which includes a fully connected layer and a layer normalization operation.

[0017] In one embodiment of the present invention, the prediction model training module integrates a multilayer perceptron classifier and a focus loss calculation unit; the output layer of the multilayer perceptron classifier does not contain a bias term to avoid probability calibration deviation; the focus loss calculation unit calculates a weighted loss based on the predicted probability and the true label, and assigns higher weights to samples with a predicted probability close to 0.5.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. This invention constructs a coronary heart disease risk assessment system with both physiological function and anatomical structure by synchronously integrating dynamic physiological time-series data acquired over a long period of time with high-resolution coronary artery structural imaging data.

[0020] 2. The dynamic physiological feature encoding module utilizes multi-scale one-dimensional convolution and bidirectional gated recurrent units to effectively capture the nonlinear coupling patterns of heart rate variability and blood pressure fluctuations at different time scales, overcoming the limitations of traditional methods that only use instantaneous or average values.

[0021] 3. The vascular structure modeling module abstracts the coronary arteries into a directed graph with attributes, and learns the relationship between local vascular lesions and global topology through a graph attention network, avoiding the neglect of vascular spatial relationships by simple image classification methods based on voxels or slices.

[0022] 4. The cross-modal feature fusion module adopts a bidirectional cross-attention mechanism, which realizes dynamic feature selection under structural constraints and structurally sensitive region enhancement under dynamic guidance, enabling the model to identify the pathological association between specific vascular stenosis regions and specific heart rate variability patterns. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall technical solution architecture of the coronary heart disease prediction model generation method based on artificial intelligence proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of cross-modal spatiotemporal alignment and feature interaction in this invention; Figure 3 This is a flowchart illustrating the multi-scale temporal feature extraction logic of dynamic physiological time-series data in this invention. Figure 4 This is a flowchart illustrating the logical flow of topological graph modeling and graph neural network representation of static vascular structure data in this invention. Figure 5 This is a logical flowchart of the multi-source data acquisition and preprocessing stage in this invention; Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the terminal and the cloud in this invention. Detailed Implementation

[0024] Please refer to the attached document. Figures 1 to 6 This invention provides a method for generating a coronary artery disease prediction model based on artificial intelligence. This method simultaneously integrates long-term continuously acquired dynamic physiological time-series data with high-resolution coronary artery structural imaging data to construct a coronary artery disease risk assessment system with dual dimensions of physiological function and anatomical structure. The method includes the following steps: S1, acquire dynamic physiological time-series data and static vascular structure data of the subject; S2, perform multi-scale temporal feature extraction on the dynamic physiological time-series data to generate dynamic physiological representations; S3, perform topological graph modeling on the static vascular structure data to generate vascular structure representation; S4, perform cross-modal spatiotemporal alignment and feature interaction between the dynamic physiological representation and the vascular structure representation to generate a fused representation; S5. Based on the fusion representation, train the coronary heart disease risk prediction model and output the probability of coronary heart disease occurrence.

[0025] In step S1, dynamic physiological time-series data and static vascular structure data of the subject are acquired. The dynamic physiological time-series data are obtained through synchronous continuous acquisition from an electrocardiogram (ECG) monitoring device and a non-invasive blood pressure monitoring device. The ECG monitoring device is configured to record standard 12-lead ECG signals, with a sampling frequency of no less than 250 Hz and a continuous acquisition time of no less than 24 hours. A heart rate variability sequence is extracted from the acquired ECG signals. This sequence includes three indicators: the standard deviation of adjacent normal heartbeat intervals, the root mean square of the difference between adjacent heartbeat intervals, and the ratio of low-frequency power to high-frequency power.

[0026] The non-invasive blood pressure monitoring device automatically measures systolic and diastolic blood pressure every 15 minutes using the oscillometric method, and simultaneously records continuous pulse waveforms. The sampling frequency is no less than 100 Hz, and the continuous acquisition time is no less than 24 hours. Blood pressure variability indices are calculated from the acquired blood pressure signals, including the standard deviation of systolic blood pressure, the coefficient of variation of diastolic blood pressure, and the amplitude of pulse pressure fluctuation. Static vascular structure data are obtained from a coronary computed tomography (CT) imaging device. This device uses multi-slice spiral CT scanning technology with a slice thickness not exceeding 0.6 mm and a reconstruction matrix of 512×512. After acquiring three-dimensional coronary artery image data through this device, the coronary artery tree structure is reconstructed, and the lumen diameter, degree of stenosis, plaque distribution location, and calcification integral of each major branch are marked. All raw data are uniformly timestamped after acquisition to ensure logical consistency between dynamic physiological data and static vascular structure data on the timeline; that is, static vascular structure data is considered a snapshot of the anatomical state during the dynamic physiological data acquisition period.

[0027] In step S2, multi-scale temporal feature extraction is performed on the dynamic physiological time-series data to generate dynamic physiological representations. First, the heart rate variability sequence and blood pressure time series are segmented according to preset time windows, each 5 minutes long with a 1-minute step, forming an overlapping sliding window sequence. Each time window contains time-series segments of five physiological indicators, constituting a five-channel one-dimensional signal input. This input is fed into a one-dimensional convolutional neural network for local feature extraction. The one-dimensional convolutional neural network contains three convolutional layers: the first layer has a kernel size of 3 and 64 output channels; the second layer has a kernel size of 5 and 128 output channels; and the third layer has a kernel size of 7 and 256 output channels. Each convolutional layer is followed by batch normalization and a modified linear unit activation function to enhance nonlinear expressive power and stabilize the training process. After three convolutional layers, each time window outputs a feature vector with a dimension of 256. Subsequently, the feature vectors of all time windows are arranged in chronological order to form a new temporal feature sequence. The sequence is input into a bidirectional gated cyclic unit for global timing modeling.

[0028] The bidirectional gated recurrent unit comprises two independent gated recurrent units, one forward and one backward, with 512 hidden units in each unit. The forward unit processes the sequence forward in time, while the backward unit processes the sequence backward in time. The hidden states of both units at each time step are concatenated to form a temporal representation that integrates bidirectional contextual information. Finally, a max-pooling operation is performed on the bidirectional hidden state sequence in the time dimension, compressing it into a fixed-length vector with a dimension of 1024. This vector represents the dynamic physiological representation, encoding the nonlinear coupled dynamic characteristics of heart rate and blood pressure at short-term (minute-level), medium-term (hour-level), and long-term (day-long) time scales, including instantaneous stress response, circadian rhythm patterns, and chronic load accumulation effects.

[0029] In step S3, topological graph modeling is performed on the static vascular structure data to generate a vascular structure representation. First, a centerline extraction algorithm is executed on the coronary artery 3D image data. This algorithm uses a fast traversal method combined with a minimum path search strategy, starting from the aortic root and tracing step by step along the main trunk to each terminal branch, generating a directed graph composed of nodes and edges. Nodes represent vascular bifurcation points or endpoints, and edges represent vascular segments connecting two nodes. Subsequently, each node is assigned an attribute vector, which contains three geometric and hemodynamic parameters: the average lumen diameter at that node, the local curvature, and the estimated wall shear stress.

[0030] Wall shear stress was estimated using Poiseuille's law combined with local vessel diameter and mean blood flow velocity. Each edge was assigned an attribute vector containing four pathological parameters: vessel segment length, mean stenosis rate, plaque load percentage, and calcification density. The mean stenosis rate was defined as (reference diameter minus minimum diameter) divided by the reference diameter, which was taken as the average of adjacent lesion-free areas. The plaque load percentage was calculated through cross-sectional analysis as the proportion of plaque area to the total vessel area. Calcification density was derived from CT values ​​and is expressed in Hounsfield units.

[0031] After assigning graph structure and attributes, the directed graph is input into a graph neural network for representation learning. The graph neural network employs a three-layer graph attention mechanism, with each layer performing message passing operations. In each layer, each node aggregates information from its first-order neighbors, with the aggregation weights dynamically calculated by the learnable attention mechanism. The attention mechanism uses a multi-head structure with eight heads, each outputting a 32-dimensional value. After concatenation and linear transformation, a 256-dimensional node embedding is obtained. After stacking the three layers, the embedding of each node incorporates structural and pathological information from its second-order neighborhood. Finally, global max pooling is performed on all node embeddings, selecting the maximum response value across all dimensions to generate a 256-dimensional vascular structure representation vector. This vector comprehensively reflects the overall topological complexity of the coronary artery system, the severity of local lesions, and the distribution of hemodynamic risks.

[0032] In step S4, the dynamic physiological representation and the vascular structure representation are aligned across modalities and their features are interacted to generate a fused representation. First, the dynamic physiological representation vector is mapped to the same embedding space as the vascular structure representation through a fully connected layer. This fully connected layer has an input dimension of 1024, an output dimension of 256, and a linear activation function. The mapped vector is called the aligned dynamic representation. Then, two cross-attention modules are constructed. The first cross-attention module uses the vascular structure representation as the query vector and the aligned dynamic representation as the key and value vectors. The query, key, and value vectors are transformed to the attention computation space through independent projection matrices, each with a projection dimension of 256. The attention weights are calculated using a scaled dot product. ; in This is a projection representing the vascular structure. and The projection of the dynamically represented image after alignment. =256, The first module is a transpose. It outputs a structure-guided dynamic feature response with a dimension of 256. The second cross-attention module uses dynamic physiological representation as the query vector and vascular structure representation as the key and value vectors, respectively. It also calculates using the above formula and outputs a dynamic-guided structural feature response with a dimension of 256. The outputs of the two cross-attention modules are concatenated along the feature dimensions to form a preliminary fusion vector of 512 dimensions. This preliminary fusion vector is then extended to 1536 dimensions via a residual connection path: first, it is extended to 1536 dimensions through a fully connected layer; then, it is added to the original concatenated vector after upsampling; finally, it is stabilized through layer normalization. The final output is a fusion representation with a dimension of 1536. This fusion representation simultaneously preserves the temporal evolution pattern of dynamic physiological signals and the spatial topological relationship of vascular structures, and establishes a fine-grained correlation between the two through a bidirectional attention mechanism. For example, stenosis in a specific vascular segment may correspond to abnormal heart rate variability at a specific time period.

[0033] In step S5, a coronary heart disease risk prediction model is trained based on the fused representation, outputting the probability of coronary heart disease occurrence. The fused representation is input into a multilayer perceptron classifier. This classifier contains three fully connected layers: the first layer has an input dimension of 1536 and an output dimension of 1000; the second layer has an input dimension of 1000 and an output dimension of 500; and the third layer has an input dimension of 500 and an output dimension of 100. Each layer uses a Gaussian error linear unit as its activation function. The final fully connected layer outputs a single scalar, which is mapped to the interval 0 to 1 using a Sigmoid function, representing the probability of coronary heart disease occurrence. The model training uses a focus loss function. Its mathematical expression is: ; in To predict the probability, if the true label is positive, then ,otherwise ; The adjustment factor is the predicted probability of coronary heart disease occurrence output by the model, mapped to the [0,1] interval by the Sigmoid function. The value is 2, which is the balance parameter. The value is set to 0.75. This loss function assigns higher weights to hard-to-classify samples with predicted probabilities close to 0.5, improving the model's ability to distinguish boundary cases. A 5-fold cross-validation strategy is used during training, randomly dividing all samples into 5 mutually exclusive subsets. Four subsets are used alternately as the training set, and the other subset is used as the validation set, repeated 5 times, and the average performance is taken. The optimizer used is AdamW, with an initial learning rate of 0.001 and a weight decay coefficient of 0.01. The learning rate decays to 1 / 10 of its original value if the validation loss does not decrease for 10 consecutive epochs. Training terminates when the validation loss shows no improvement for 20 consecutive epochs. The final model achieves an area under the receiver operating characteristic (AUC) of 0.92 on the independent test set, and its detection sensitivity for early asymptomatic coronary heart disease patients is more than 28% higher than traditional methods.

[0034] The AI-based coronary artery disease prediction model generation system includes a multi-source data acquisition module, a dynamic physiological feature encoding module, a vascular structure modeling module, a cross-modal feature fusion module, and a prediction model training module. The multi-source data acquisition module includes an electrocardiogram (ECG) signal acquisition unit, a blood pressure signal acquisition unit, and a coronary artery imaging acquisition unit. The ECG signal acquisition unit is configured to continuously record standard 12-lead ECG signals at a sampling frequency of no less than 250 Hz; the blood pressure signal acquisition unit is configured to automatically measure blood pressure every 15 minutes using oscillometric methods and simultaneously record continuous pulse waveforms at a sampling frequency of no less than 100 Hz; the coronary artery imaging acquisition unit is configured to use a multi-slice spiral computed tomography (CT) scanner with a slice thickness not exceeding 0.6 mm and a reconstruction matrix of 512×512.

[0035] The dynamic physiological feature encoding module integrates a one-dimensional convolutional neural network and a bidirectional gated recurrent unit. The one-dimensional convolutional neural network has 5 input channels, corresponding to three heart rate variability indicators and two indicators: systolic and diastolic blood pressure. The forward and backward hidden states of the bidirectional gated recurrent unit are concatenated in the time dimension and compressed into a fixed-length vector through time pooling. The vascular structure modeling module integrates a centerline extraction algorithm unit and a graph neural network unit. The centerline extraction algorithm unit uses a fast traversal method combined with minimum path search to extract the centerlines of the main trunk and branches from the three-dimensional coronary artery images. The graph neural network unit is configured to receive attribute tensors of nodes and edges and execute a message passing mechanism, fusing its own features and weighted features of neighboring nodes when updating node embeddings at each layer.

[0036] The cross-modal feature fusion module integrates two cross-attention sub-modules and a feature concatenation unit. Each cross-attention sub-module contains a query projection matrix, a key projection matrix, and a value projection matrix, all with a projection dimension of 256. The feature concatenation unit concatenates the outputs of the two sub-modules along the feature dimensions and connects them to a residual connection path, which includes a fully connected layer and layer normalization operations. The prediction model training module integrates a multilayer perceptron classifier and a focus loss calculation unit. The output layer of the multilayer perceptron classifier does not contain a bias term to avoid probability calibration bias. The focus loss calculation unit calculates a weighted loss based on the predicted probability and the true label, assigning higher weights to samples with predicted probabilities close to 0.5.

[0037] During system operation, the multi-source data acquisition module first completes data collection and preprocessing, outputting standardized dynamic physiological time-series data and static vascular structure data. The dynamic physiological feature encoding module receives the dynamic physiological time-series data, performs multi-scale time-series feature extraction, and outputs dynamic physiological representations. The vascular structure mapping module receives the static vascular structure data, performs topological graph modeling and graph neural network representation, and outputs vascular structure representations. The cross-modal feature fusion module receives the above two representations, performs spatiotemporal alignment and bidirectional cross-attention interaction, and outputs fused representations. The prediction model training module receives the fused representations, performs classification training, and outputs probabilities. The entire system can be deployed on local medical terminals or through a cloud server for distributed training and inference. During data transmission, an encryption protocol is used to protect patient privacy, and model updates are performed through a federated learning mechanism to aggregate parameters without sharing the original data. The system's output probability of coronary heart disease can be directly used by clinicians to assist in developing personalized intervention strategies, such as lifestyle modifications, drug treatment, or further invasive examinations.

Claims

1. A method for generating a coronary heart disease prediction model based on artificial intelligence, characterized in that, include: Acquire dynamic physiological time-series data and static vascular structure data of the subjects being examined; Multi-scale temporal feature extraction is performed on the dynamic physiological time-series data to generate dynamic physiological representations; Topological graph modeling is performed on the static vascular structure data to generate vascular structure representation; The dynamic physiological representation and the vascular structure representation are aligned across modalities and interact with each other in a spatiotemporal manner to generate a fused representation; The coronary heart disease risk prediction model is trained based on the fusion representation and outputs the probability of coronary heart disease occurrence.

2. The method for generating a coronary heart disease prediction model based on artificial intelligence according to claim 1, characterized in that, The acquisition of dynamic physiological time-series data and static vascular structure data of the subject includes: ECG signals are continuously collected for no less than 24 hours using ECG monitoring equipment, and heart rate variability sequences are extracted from them. The heart rate variability sequences include the standard deviation of adjacent normal heartbeat intervals, the root mean square of the difference between adjacent heartbeat intervals, and the ratio of low-frequency power to high-frequency power. The systolic and diastolic blood pressure time series were collected synchronously for no less than 24 hours using a non-invasive blood pressure monitoring device, and the blood pressure variability index was calculated. The blood pressure variability index includes the standard deviation of systolic blood pressure, the coefficient of variation of diastolic blood pressure, and the pulse pressure fluctuation amplitude. Three-dimensional images of the coronary arteries are obtained using coronary computed tomography (CT) imaging equipment to reconstruct the tree-like structure of the coronary arteries and mark the lumen diameter, degree of stenosis, plaque distribution location, and calcification score of each major branch.

3. The method for generating a coronary heart disease prediction model based on artificial intelligence according to claim 2, characterized in that, Multi-scale temporal feature extraction is performed on the dynamic physiological time-series data to generate dynamic physiological representations, including: The heart rate variability series and blood pressure time series were divided into multiple subsequences according to a preset time window, with each time window being 5 minutes long and the step size being 1 minute. Each subsequence is input into a one-dimensional convolutional neural network, which contains three convolutional layers with kernel sizes of 3, 5, and 7, and channel numbers of 64, 128, and 256, respectively. Each layer is followed by batch normalization and modified linear unit activation functions. The convolutional outputs of each time window are used for temporal modeling through bidirectional gated recurrent units, with a hidden unit count of 512. The final output is a dynamic physiological representation vector with a dimension of 1024, which encodes the nonlinear coupling dynamic characteristics of heart rate and blood pressure at short, medium and long time scales.

4. The method for generating a coronary heart disease prediction model based on artificial intelligence according to claim 3, characterized in that, Topological graph modeling is performed on the static vascular structure data to generate vascular structure representations, including: The coronary artery 3D image data is processed by a centerline extraction algorithm to generate a directed graph consisting of nodes and edges, where nodes represent blood vessel bifurcation points or endpoints, and edges represent blood vessel segments. Each node is assigned an attribute vector, which includes the average lumen diameter, local curvature, and estimated wall shear stress at that node. Each edge is assigned an attribute vector, which includes the length of the vessel segment, the average stenosis rate, the plaque burden percentage, and the calcification density. The directed graph is input into a graph neural network, which employs a three-layer graph attention mechanism. Each layer calculates the attention weights between nodes when aggregating neighborhood information. The number of attention heads is 8, and the output dimension is 256. After global max pooling, a vascular structure representation vector with dimension 256 is generated.

5. The method for generating a coronary heart disease prediction model based on artificial intelligence according to claim 4, characterized in that, The dynamic physiological representation and the vascular structure representation are aligned across modalities and their features are interacted to generate a fused representation, including: The dynamic physiological representation vector is mapped to the same embedding space as the vascular structure representation through a fully connected layer to obtain the aligned dynamic representation. A cross-attention module is constructed, using the vascular structure representation as the query vector and the aligned dynamic representation as the key and value vector, to calculate the structure-guided dynamic feature response; Simultaneously, another cross-attention module is constructed, using dynamic physiological representation as the query vector and vascular structure representation as the key and value vector, to calculate the dynamically guided structural feature response; The outputs of the two cross-attention modules are concatenated and then normalized by residual connections to generate a fused representation with a dimension of 1536.

6. The method for generating a coronary heart disease prediction model based on artificial intelligence according to claim 5, characterized in that, The dynamic physiological representation vector is mapped to the same embedding space as the vascular structure representation through a fully connected layer to obtain the aligned dynamic representation, including: The fully connected layer has an input dimension of 1024, an output dimension of 256, and a linear activation function.

7. The method for generating a coronary heart disease prediction model based on artificial intelligence according to claim 5, characterized in that, A cross-attention module is constructed, using vascular structure representation as the query vector and aligned dynamic representation as the key-value vector, to compute structure-guided dynamic feature responses, including: The query, key, and value vectors are each transformed into the attention computation space through independent projection matrices, with a projection dimension of 256 for each. Attention weights are calculated using a scaled dot product attention mechanism, and a structure-guided dynamic feature response with a dimension of 256 is output.

8. The method for generating a coronary heart disease prediction model based on artificial intelligence according to claim 5, characterized in that, Construct another cross-attention module, using dynamic physiological representations as query vectors and vascular structure representations as key-value vectors, to compute dynamically guided structural feature responses, including: The query, key, and value vectors are each transformed into the attention computation space through independent projection matrices, with a projection dimension of 256 for each. Attention weights are calculated using a scaled dot product attention mechanism, and a dynamically guided structural feature response with a dimension of 256 is output.

9. The method for generating a coronary heart disease prediction model based on artificial intelligence according to claim 5, characterized in that, The outputs of the two cross-attention modules are concatenated, and a fused representation of dimension 1536 is generated through residual connections and layer normalization, including: The outputs of the two cross-attention modules are concatenated along the feature dimension to form a preliminary 512-dimensional fusion vector. The initial fused vector is extended to 1536 dimensions through a fully connected layer and added to the upsampled original concatenated vector; Perform layer normalization on the addition result and output the fusion representation.

10. The method for generating a coronary heart disease prediction model based on artificial intelligence according to claim 1, characterized in that, Based on the aforementioned fusion representation, a coronary heart disease risk prediction model is trained, outputting the probability of coronary heart disease occurrence, including: The fused representation is input into a multilayer perceptron classifier, which contains three fully connected layers with 1000, 500, and 100 hidden units, respectively, and the activation function is Gaussian error linear unit. The last layer outputs a single scalar, which is mapped to the probability of coronary heart disease by the sigmoid function. Model training uses a focus loss function and an adjustment factor. The value is 2, which is the balance parameter. The value is 0.75; Five-fold cross-validation was used during training, the optimizer was AdamW, the initial learning rate was 0.001, and the weight decay coefficient was 0.01.

Images