Cardiovascular disease dynamic risk trajectory prediction method and system
By constructing a multimodal deep learning architecture and a large language model, the problems of multimodal data fusion and user interaction in dynamic risk prediction of cardiovascular diseases are solved, achieving high-precision, interactive, and personalized risk prediction, and supporting individualized management strategies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUDAN UNIVERSITY
- Filing Date
- 2026-01-19
- Publication Date
- 2026-04-28
AI Technical Summary
Existing technologies cannot effectively predict the dynamic risk trajectory of cardiovascular diseases, lack multimodal data fusion and intelligent user interaction capabilities, resulting in difficulties in identifying disease progression patterns, a lack of biological rationality in prediction results, and an inability to support personalized decision-making.
We construct an end-to-end multimodal deep learning architecture, combining Transformer and variational autoencoder technologies to achieve cross-modal fusion of electrocardiograms and electronic health records, generate evolutionary digital risk fingerprints, and realize natural language interaction through a large language model to support personalized risk prediction.
It improves the accuracy and robustness of cardiovascular disease risk prediction, identifies key nodes in disease progression, supports personalized management strategies, and enables efficient human-machine collaborative decision-making.
Smart Images
Figure CN121938633A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method and system for predicting dynamic risk trajectories of cardiovascular diseases. Background Technology
[0002] Cardiovascular disease is the leading cause of death worldwide, encompassing various types such as coronary artery disease, heart failure, arrhythmia, and cardiomyopathy. It is characterized by high mortality and high readmission rates. The onset of cardiovascular disease is difficult to predict, and its progression is complex and variable, exhibiting non-linear evolutionary characteristics. The development of pathophysiological stages can span from months to years, posing a significant challenge to early risk identification.
[0003] Traditional risk scoring methods (such as the Framingham Risk Score, GRACE Score, and PCP-HF Score) and existing machine learning prediction models can only achieve cross-sectional risk stratification, treating disease risk as a static attribute of an individual rather than a dynamic evolutionary process. These methods essentially compress complex disease evolution into a single probability estimate, failing to identify key progression points and different risk evolution patterns. Furthermore, most studies are limited to a single data modality, failing to integrate multi-source heterogeneous data such as electrocardiograms and electronic health records, thus failing to depict a complete picture of the disease. Moreover, artificial intelligence models generally suffer from the "black box" problem; users cannot understand the biological rationale behind the predictions, limiting the models' application. Most importantly, existing models lack the ability to intelligently interact with users, failing to utilize user experience for dynamic decision optimization, and serving only as passive prediction tools.
[0004] Therefore, there is an urgent need to develop a new predictive system that can predict the dynamic trajectory of personalized cardiovascular disease risk, support multimodal data fusion, and interact intelligently with users, in order to meet the actual needs of users for decision support. Summary of the Invention
[0005] The purpose of this invention is to propose a method and system for predicting dynamic risk trajectories of cardiovascular diseases. This invention constructs an end-to-end multimodal deep learning architecture while maximizing prediction accuracy. Based on Transformer and variational autoencoder technologies, it performs cross-modal fusion of electrocardiograms and electronic health records, fully extracting complementary information from different data sources. Risk trajectory modeling is achieved based on discrete-time risk prediction. Subsequently, an evolutionary digital risk fingerprint is generated through queue-specific time-point calibration to identify typical risk trajectory types. Finally, a large language model enables natural language interaction and user decision support, thereby achieving high-precision, interactive, and personalized dynamic prediction of cardiovascular disease risk.
[0006] The present invention proposes a method for predicting dynamic risk trajectories of cardiovascular diseases, the specific steps of which are as follows: (1) Preprocessing and feature space mapping of multimodal data, that is, deploying data preprocessing modules for electrocardiograms and electronic health records of different individuals and obtaining initial features: for electrocardiogram preprocessing, the electrocardiogram data matrix E is obtained, and for electronic health records, the text data matrix R is obtained; (2) Dataset partitioning: For the electrocardiogram data matrix E and text data matrix R obtained in step (1), the data is partitioned into training set {E, R, y} according to their natural chronological order and a preset ratio. train and test set {E, R, y} test Where: y is the cardiovascular disease event label; (3) Construct a multimodal deep learning model for multimodal feature representation of cardiovascular disease risk prediction. The multimodal deep learning model includes an electrocardiogram encoding module, a text encoding module, a cross-modal fusion module, and a discrete-time risk prediction module. The ECG encoding module is built based on a Transformer encoder block, configured with an attention mechanism and a feedforward network, for modeling the temporal dependencies of the ECG and extracting features. In the pre-training stage, a contrastive learning mechanism and a frequency band mapping auxiliary task are used to perform self-supervised pre-training on the ECG encoding module. The loss function in the pre-training stage includes contrastive learning loss and frequency band mapping reconstruction loss. In the main training stage, the ECG data matrix E obtained in step (1) is input into the ECG encoding module to obtain the ECG feature vector. , where d is the uniform feature dimension and N is the number of samples; The text encoding module, based on a text model, extracts semantic features from electronic health records through a multi-layer architecture of marker embedding, fragment embedding, and position embedding. The text data matrix R obtained in step (1) is input into the text encoding module, and after dimensionality reduction by a projection layer, a text feature vector is obtained. ; The cross-modal fusion module is built based on the principle of variational autoencoder. Through encoder network, reparameterization technology and decoder network, it realizes joint latent representation learning of electrocardiogram feature vectors and text feature vectors to obtain fused multimodal feature vectors. The discrete-time risk prediction module consists of multiple activated neurons, each corresponding to a time interval within a preset time period. The prediction module is used to analyze the multimodal feature vector F. m Perform regression calculations to generate conditional risk probabilities for multiple time points. , where K is the number of time points; (4) Constructing the multimodal risk prediction loss function in the main training phase: The multimodal risk prediction loss function is obtained by weighted combination of three loss functions: discrete-time conditional likelihood loss, multimodal contrastive learning loss, and cumulative risk loss; (5) Constructing a time calibration module to generate an evolutionary digital risk fingerprint: Based on the probability calibration method, a specific time point calibrator is trained using some training data, and a personalized probability calibration function is established to convert the conditional risk probability h obtained in step (3) into the calibrated cumulative risk probability H; an interpolation technique is used to construct a smooth risk trajectory curve to generate an evolutionary digital risk fingerprint. (6) Model training, testing and evaluation: The training set {E, R, y} obtained in step (2) is used for training, testing and evaluation. train The input is fed into the multimodal deep learning model constructed in step (3) for training. The multimodal risk prediction loss function constructed in step (4) is used for iterative optimization. The optimizer is used to update the parameters. Cross-validation and early stopping mechanisms are adopted to obtain the optimized cardiovascular disease risk prediction model. In the testing phase, the training set {E, R, y} is used. test Input the optimized cardiovascular disease risk prediction model, output the risk prediction probability through forward propagation, generate an individualized risk trajectory according to the time calibration module in step (5), and identify the risk trajectory type; finally, calculate the evaluation index to verify the accuracy of the model in dynamic risk prediction. (7) Construct an intelligent interaction layer based on a large language model: The intelligent interaction layer uses a large language model as the central inference engine and realizes information interaction between the large language model and the multimodal deep learning model through a standardized communication protocol, so as to realize the dynamic risk trajectory prediction of individuals and interactive user decision support.
[0007] In this invention, in the electrocardiogram encoding module described in step (3), the contrastive learning mechanism maximizes the consistency between the original electrocardiogram and the enhanced view, and minimizes the consistency between the original electrocardiogram and other electrocardiograms and their enhanced views; the frequency band mapping auxiliary task decomposes the electrocardiogram data into multiple frequency bands and reconstructs them.
[0008] In this invention, in the multimodal risk prediction loss function described in step (4), the discrete-time conditional likelihood loss penalizes the difference between the predicted conditional risk probability and the actual occurrence of the event; the multimodal contrastive learning loss maximizes the cross-modal feature similarity of the same sample and minimizes the cross-modal feature similarity of different samples; and the cumulative risk loss penalizes the difference between the cumulative risk probability and the actual cumulative event rate.
[0009] In this invention, the large language model mentioned in step (7) is responsible for the natural language interpretation of the prediction results, user dialogue management, and multi-round interactive decision-making; dynamically calls the multimodal deep learning model to recalibrate the prediction results based on user feedback; and automatically generates a structured report containing individual information, risk assessment, trajectory prediction, reasoning basis, and personalized management strategies.
[0010] In this invention, by replacing the electrocardiogram encoding module with other physiological information encoding modules such as pulse wave and blood oxygen, the method can be extended to dynamic risk prediction of other chronic diseases such as hypertension and diabetes.
[0011] In this invention, the dynamic risk prediction is the risk of new cardiovascular events within a preset time period, including one or more of heart failure, myocardial infarction, or arrhythmia; for heart failure, it also includes subtype or functional classification prediction.
[0012] In this invention, the electrocardiogram in step (1) is one of single-lead, three-lead, or twelve-lead, supporting resting or dynamic electrocardiograms; the structured extraction of the electronic health record uses the natural language understanding ability of a large language model, and the preset individual information extracted includes at least one of demographic information, personal history, comorbidities, treatment history, present illness history, or diagnostic history.
[0013] In this invention, the time calibration module in step (5) performs probability calibration based on Platt scaling or isotonic regression methods; the risk trajectory type includes one of high-risk rapid progression, medium-risk volatile, medium-risk stable, and low-risk slow progression.
[0014] The prediction system for dynamic risk trajectory prediction of cardiovascular diseases proposed in this invention includes: a multimodal perception layer, an intelligent inference layer, an interaction protocol layer, and a personalized output layer; wherein: The multimodal sensing layer is used to receive electrocardiograms and electronic health records, resample and normalize the electrocardiogram data, and extract the structure from the electronic health records. The intelligent inference layer includes a large language model inference engine and a multimodal deep learning model, wherein the multimodal deep learning model includes an electrocardiogram encoding module, a text encoding module, a cross-modal fusion module, and a discrete-time risk prediction module; The interaction protocol layer is used to achieve standardized communication and state synchronization between large language models and multimodal deep learning models, and to save the session context through a state machine-driven interactive workflow. A personalized output layer is used to generate evolving digital risk fingerprints, identify risk trajectory types, and generate structured reports.
[0015] The present invention has the following beneficial effects: 1. This invention addresses the modal heterogeneity of electrocardiograms and electronic health records by using variational autoencoders to achieve joint representation learning of cross-modal features, effectively extracting complementary information from different data sources and significantly improving the accuracy and robustness of cardiovascular disease risk prediction. 2. This invention generates conditional risk probabilities at multiple time points through a discrete-time risk prediction module, constructs a dynamic risk evolution trajectory, and can identify key nodes in disease progression and different risk evolution patterns, overcoming the limitations of traditional static risk assessment and providing a scientific basis for the formulation of personalized long-term management strategies. 3. This invention employs an electrocardiogram encoding module based on Transformer and a text encoding module based on a pre-trained text model. It captures the long-term temporal dependence of electrocardiograms and the semantic features of electronic health records through an attention mechanism, thereby achieving end-to-end multi-level automatic feature learning and effectively avoiding the subjectivity and incompleteness of manual feature engineering. 4. This invention utilizes a time calibration module to generate an evolutionary digital risk fingerprint, identify typical risk trajectory types, and achieve personalized and precise stratification. It calibrates the risk probability to the actual event occurrence rate, ensuring the accuracy of individual risk assessment and the reliability of user decisions. 5. This invention combines a large language model to achieve natural language interaction, dynamic correction of prediction results, and automatic generation of structured reports. It supports users in supplementing individual information through dialogue, questioning prediction results, and enabling users to understand the model's reasoning basis, thereby achieving efficient human-computer collaborative decision-making. 6. This invention provides an interactive method for dynamic risk prediction of cardiovascular diseases, which can be applied to scenarios such as early warning of various diseases such as heart failure and myocardial infarction, screening of high-risk individuals, and formulation of personalized management strategies. It has important application value in the prevention and treatment of cardiovascular diseases and in the field of multimodal medical artificial intelligence, and can be extended to dynamic risk prediction and management of chronic diseases based on other physiological information. Attached Figure Description
[0016] Figure 1 The flowchart of dynamic risk prediction for heart failure based on electrocardiogram and electronic medical record is shown in the embodiment of the present invention.
[0017] Figure 2 The following diagram illustrates an example of multimodal data input in an embodiment of the present invention, including (a) the original twelve-lead electrocardiogram (including leads I, II, III, aVR, aVL, aVF, and V1-V6), and (b) the structured text extracted from the electronic health record by the large language model.
[0018] Figure 3 The training process and architecture diagram of the multimodal deep learning model in the embodiments of the present invention are shown, including the complete architecture of the electrocardiogram encoder pre-training stage and the main training stage.
[0019] Figure 4 This invention demonstrates four typical risk trajectory types identified based on hierarchical clustering algorithm in this embodiment. The horizontal axis represents the prediction time (months), and the vertical axis represents the cumulative risk probability. These include (a) high-risk rapid progression type, (b) medium-risk volatile type, (c) medium-risk stable type, and (d) low-risk slow progression type.
[0020] Figure 5 This paper illustrates the intelligent interaction layer architecture based on a large language model in an embodiment of the present invention, including the standardized communication protocol and interaction workflow between the large language model and the multimodal deep learning model.
[0021] Figure 6 The diagram illustrates the structure of the prediction system used in the dynamic risk trajectory prediction of cardiovascular diseases according to this invention, which includes a multimodal perception layer, an intelligent inference layer, an interaction protocol layer, and a personalized output layer. Detailed Implementation
[0022] The method and its application of the present invention will be further described below with reference to the accompanying drawings and embodiments. These embodiments do not limit the present invention; structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are all included within the scope of protection of the present invention.
[0023] Example 1: Using 7,123 individuals from real-world scenarios as the research cohort, and based on 15,212 electrocardiogram (ECG) and electronic health record samples, the cardiovascular disease dynamic risk trajectory prediction method of this invention was applied to predict the risk of new-onset heart failure. The ECGs were 12-lead ECGs (10 seconds, 500Hz sampling rate), and the electronic health records included demographic information, personal history, chronic disease history, treatment history, present illness history, and diagnostic history. The prediction target was the risk of new-onset heart failure events within 24 months. Using the cardiovascular disease dynamic risk trajectory prediction method of this invention, this example predicts the dynamic risk of heart failure based on ECGs and electronic medical records. The flowchart is as follows: Figure 1 As shown. The specific steps are as follows: (1) Examples of multimodal data used in this embodiment are as follows: Figure 2 As shown, Figure 2 (a) is the original twelve-lead electrocardiogram (including leads I, II, III, aVR, aVL, aVF, and V1-V6). Figure 2(b) Structured text extracted from electronic health records by the Dawei language model. Raw electrocardiogram (ECG) signals and raw electronic health records were acquired. For ECGs, the ECG signals with a sampling rate of 500 Hz were first resampled to 128 Hz and normalized to 1,280 sampling points. Then, z-score normalization was used to eliminate amplitude differences between individuals. Finally, the 12-lead signals were mapped to a 350-dimensional feature space through a linear mapping layer to obtain an ECG data matrix with dimensions [15, 212, 1280, 350].
[0024] For electronic health records, the DeepSeek large language model is used to extract demographic information (age and gender), personal history (smoking and drinking history), comorbidities, treatment history, present illness history, and diagnosis history from unstructured text and then structure it. The extracted structured text is then tokenized using the BERT tokenizer and normalized to 350 token sequences, resulting in a text data matrix with dimensions [15, 212, 350].
[0025] (2) The ECG data matrix and text data matrix obtained in step (1) were divided into datasets. The 15,212 samples from 7,123 individuals were divided into training and test sets according to the inter-individual paradigm in a ratio of 80%:20%. The training set contained 12,169 samples from 5,698 individuals, including 1,344 cases of heart failure events in 24 months; the test set contained 3,043 samples from 1,425 individuals, including 336 cases of heart failure events in 24 months.
[0026] (3) such as Figure 3 As shown, a multimodal deep learning model is constructed, which includes four main parts: an electrocardiogram encoding module, a text encoding module, a cross-modal fusion module, and a risk prediction module. Model training is divided into two steps: a pre-training phase and a main training phase. During the pre-training phase, the ECG encoding module adopted a Transformer architecture with 5 attention heads and was pre-trained on the PTB-XL dataset (21,837 individuals) using contrastive learning loss. and frequency band mapping reconstruction loss Combinations: Learn effective representations of electrocardiograms.
[0027] During the main training phase, the ECG encoding module outputs feature vectors. The text encoding module uses the BERT model, which reduces the 768-dimensional features to 350 dimensions through a projection layer, outputting a feature vector. The cross-modal fusion module is based on the principle of variational autoencoder, which integrates ECG features F... e and text features F rThe latent representations are mapped to the mean μ and variance σ² of the latent space, respectively. They are then sampled using a reparameterization technique to obtain the latent representations, which are concatenated and mapped to the fused features. The risk prediction module uses 12 sigmoid-activated neurons to correspond to 12 time intervals within 24 months (each interval is 2 months), and outputs the conditional risk probability h∈ℝ for each time interval. N×12 .
[0028] (4) Construct the multimodal risk prediction loss function for the main training phase, including discrete-time conditional likelihood loss. Multimodal contrastive learning loss and cumulative risk loss The discrete-time conditional likelihood loss penalizes the difference between the predicted conditional risk probability and the actual occurrence of events; the multimodal contrastive learning loss promotes cross-modal feature alignment and enhances multimodal data fusion capabilities; the cumulative risk loss penalizes the difference between the cumulative risk probability and the actual cumulative event rate, ensuring the accuracy of long-term risk estimation. The multimodal risk prediction loss function is obtained by weighting and combining the three loss functions: ,in This is the balance coefficient.
[0029] (5) Construct a time calibration module for generating evolutionary digital risk fingerprints: Train a time point calibrator using 5% of the training data, and use the Platt scaling method to convert the conditional risk probability h into the cumulative risk probability H. Use cubic spline interpolation to smooth the discrete risk probability points and generate continuous individualized risk trajectory curves.
[0030] (6) Model Training, Testing, and Evaluation: The training set obtained in step (2) is input into the multimodal deep learning model constructed in step (3) for training, and the loss function constructed in step (4) is used for iterative optimization. The Adam optimizer is used in the pre-training stage of the ECG encoder, with a learning rate of 1×10⁻⁶. -4 The batch size is 64. Pre-training data is divided into training and validation sets in a 9:1 ratio, with a maximum of 100 training epochs. An early stopping mechanism is implemented when the validation loss fails to improve for three consecutive epochs. During the main training phase, the multimodal deep learning model employs an 8-fold cross-validation strategy on the training set. For each fold, 30 epochs are set, and the AdamW optimizer is used for parameter updates with a learning rate of 1×10⁻⁶. -5 The weights are decayed to 0.005, and the optimal model weights are saved based on the overall C-index on the validation set.
[0031] During the testing phase, the test set is input into the optimal multimodal deep learning model, and the risk prediction probability is output through forward propagation. Individualized risk trajectories (evolving digital risk fingerprints) are generated according to the time calibration module in step (5); hierarchical clustering algorithm (Ward connection criterion) is used to perform cluster analysis on the 12-dimensional risk trajectory vectors, identifying four typical risk evolution patterns, such as... Figure 4 As shown, it includes (a) high-risk rapid progress type, (b) medium-risk volatile type, (c) medium-risk stable type and (d) low-risk slow progress type.
[0032] Table 1 shows the accuracy of multimodal deep learning models in dynamic risk prediction: the C-index evaluates the model's discriminative ability, AUC evaluates the classification performance, and Brier score evaluates the accuracy of the prediction.
[0033] Table 1. Predictive performance evaluation of multimodal deep learning models on the test set. .
[0034] (7) For example Figure 5 As shown, an intelligent interaction layer is built based on the DeepSeek large language model, connecting multimodal deep learning models through the Model Context Protocol (MCP). The interaction interface adopts a state machine-driven multi-turn dialogue mechanism, supporting the supplementation of missing information, modification of prediction conditions, and discussion via natural language. When changes in individual information are detected, risk reassessment and dynamic calibration are automatically triggered.
[0035] (8) In this embodiment, the system acquires an individual's electrocardiogram and electronic health record. After standardized preprocessing, the data is first input into a multimodal deep learning model for feature extraction and fusion. Then, an individualized risk trajectory curve and type are generated. The system provides reasoning basis, prediction results, and suggestions through an intelligent interaction layer. Finally, the system automatically generates a structured report, which includes basic individual information, risk assessment results, trajectory prediction curve, reasoning basis, and management strategy suggestions. The report supports DOCX output format.
[0036] This embodiment demonstrates that the method of the present invention fully integrates the physiological information of electrocardiogram and the textual information of electronic health record, and provides a scientific basis and personalized management solution for user decision support through dynamic risk trajectory prediction and intelligent interaction.
Claims
1. A method for predicting dynamic risk trajectories of cardiovascular diseases, characterized in that, The specific steps are as follows: (1) Preprocessing and feature space mapping of multimodal data, that is, deploying data preprocessing modules for electrocardiograms and electronic health records of different individuals and obtaining initial features: for electrocardiogram preprocessing, the electrocardiogram data matrix E is obtained, and for electronic health records, the text data matrix R is obtained; (2) Dataset partitioning: For the electrocardiogram data matrix E and text data matrix R obtained in step (1), the data is partitioned into training set {E, R, y} according to their natural chronological order and a preset ratio. train and the test set {E, R, y} test Where: y is the cardiovascular disease event label; (3) Construct a multimodal deep learning model for multimodal feature representation of cardiovascular disease risk prediction. The multimodal deep learning model includes an electrocardiogram encoding module, a text encoding module, a cross-modal fusion module, and a discrete-time risk prediction module. The ECG encoding module is built based on a Transformer encoder block, configured with an attention mechanism and a feedforward network, for modeling the temporal dependencies of the ECG and extracting features. In the pre-training stage, a contrastive learning mechanism and a frequency band mapping auxiliary task are used to perform self-supervised pre-training on the ECG encoding module. The loss function in the pre-training stage includes contrastive learning loss and frequency band mapping reconstruction loss. In the main training stage, the ECG data matrix E obtained in step (1) is input into the ECG encoding module to obtain the ECG feature vector. , where d is the uniform feature dimension and N is the number of samples; The text encoding module, based on a text model, extracts semantic features from electronic health records through a multi-layer architecture of marker embedding, fragment embedding, and position embedding. The text data matrix R obtained in step (1) is input into the text encoding module, and after dimensionality reduction by a projection layer, a text feature vector is obtained. ; The cross-modal fusion module is built based on the principle of variational autoencoder. Through encoder network, reparameterization technology and decoder network, it realizes joint latent representation learning of electrocardiogram feature vectors and text feature vectors to obtain fused multimodal feature vectors. ; The discrete-time risk prediction module consists of multiple activated neurons, each corresponding to a time interval within a preset time period. The prediction module is used to analyze the multimodal feature vector F. m Perform regression calculations to generate conditional risk probabilities for multiple time points. , where K is the number of time points; (4) Constructing the multimodal risk prediction loss function in the main training phase: The multimodal risk prediction loss function is obtained by weighted combination of three loss functions: discrete-time conditional likelihood loss, multimodal contrastive learning loss, and cumulative risk loss; (5) Constructing a time calibration module to generate an evolutionary digital risk fingerprint: Based on the probability calibration method, a specific time point calibrator is trained using some training data, and a personalized probability calibration function is established to convert the conditional risk probability h obtained in step (3) into the calibrated cumulative risk probability H; an interpolation technique is used to construct a smooth risk trajectory curve to generate an evolutionary digital risk fingerprint. (6) Model training, testing and evaluation: The training set {E, R, y} obtained in step (2) is used for training, testing and evaluation. train The input is fed into the multimodal deep learning model constructed in step (3) for training. The multimodal risk prediction loss function constructed in step (4) is used for iterative optimization. The optimizer is used to update the parameters. Cross-validation and early stopping mechanisms are adopted to obtain the optimized cardiovascular disease risk prediction model. In the testing phase, the training set {E, R, y} is used. test Input the optimized cardiovascular disease risk prediction model, output the risk prediction probability through forward propagation, generate an individualized risk trajectory according to the time calibration module in step (5), and identify the risk trajectory type; finally, calculate the evaluation index to verify the accuracy of the model in dynamic risk prediction. (7) Construct an intelligent interaction layer based on a large language model: The intelligent interaction layer uses a large language model as the central inference engine and realizes information interaction between the large language model and the multimodal deep learning model through a standardized communication protocol, so as to realize the dynamic risk trajectory prediction of individuals and interactive user decision support.
2. The method according to claim 1, characterized in that, In the ECG encoding module described in step (3), the contrastive learning mechanism maximizes the consistency between the original ECG and the enhanced view, and minimizes the consistency between the original ECG and other ECGs and their enhanced views; the frequency band mapping auxiliary task decomposes the ECG data into multiple frequency bands and reconstructs them.
3. The method according to claim 1, characterized in that, In step (4), the discrete-time conditional likelihood loss penalizes the difference between the predicted conditional risk probability and the actual occurrence of the event; the multimodal contrastive learning loss maximizes the cross-modal feature similarity of the same sample and minimizes the cross-modal feature similarity of different samples. Cumulative risk loss penalty is the difference between the cumulative risk probability and the actual cumulative event rate.
4. The method according to claim 1, characterized in that, The large language model described in step (7) is responsible for the natural language interpretation of the prediction results, user dialogue management, and multi-round interactive decision-making; dynamically calling the multimodal deep learning model to recalibrate the prediction results based on user feedback; and automatically generating a structured report containing individual information, risk assessment, trajectory prediction, reasoning basis, and personalized management strategies.
5. The method according to claim 1, characterized in that, By replacing the electrocardiogram encoding module with a pulse wave and blood oxygen physiological information encoding module, the method can be extended to dynamic risk prediction of hypertension, diabetes and other chronic diseases.
6. The method according to claim 1, characterized in that, The dynamic risk prediction is the risk of new cardiovascular events within a preset time period, including one or more of heart failure, myocardial infarction, or arrhythmia; for heart failure, it also includes subtype or functional classification prediction.
7. The method according to claim 1, characterized in that, The electrocardiogram mentioned in step (1) is one of single-lead, three-lead, or twelve-lead, supporting resting or dynamic electrocardiograms; the structured extraction of the electronic health record uses the natural language understanding ability of the large language model, and the preset individual information extracted includes at least one of demographic information, personal history, comorbidities, treatment history, present illness history, or diagnosis history.
8. The method according to claim 1, characterized in that, The time calibration module in step (5) performs probability calibration based on Platt scaling or isotonic regression methods; the risk trajectory type includes one of the following: high-risk rapid progression, medium-risk volatile, medium-risk stable, and low-risk slow progression.
9. A prediction system for predicting dynamic risk trajectories of cardiovascular diseases, characterized in that, include: The system comprises a multimodal perception layer, an intelligent inference layer, an interaction protocol layer, and a personalized output layer; among which: The multimodal sensing layer is used to receive electrocardiograms and electronic health records, resample and normalize the electrocardiogram data, and extract the structure from the electronic health records. The intelligent inference layer includes a large language model inference engine and a multimodal deep learning model, wherein the multimodal deep learning model includes an electrocardiogram encoding module, a text encoding module, a cross-modal fusion module, and a discrete-time risk prediction module; The interaction protocol layer is used to achieve standardized communication and state synchronization between large language models and multimodal deep learning models, and to save the session context through a state machine-driven interactive workflow. A personalized output layer is used to generate evolving digital risk fingerprints, identify risk trajectory types, and generate structured reports.