A cardiac status risk prediction model training system

Abstract

The application provides a heart state risk prediction model training system, which comprises a data acquisition and preprocessing module, a multi-modal data adaptive fusion module, an attention mechanism-based deep prediction network and a risk prediction and visualization output module. The data acquisition and preprocessing module is used for acquiring multi-modal physiological signal data and preprocessing the data. The multi-modal data adaptive fusion module is used for receiving the preprocessed multi-modal physiological signal data, adaptively weighting and fusing the features of different modalities through a gating fusion mechanism and outputting a fused feature vector. The attention mechanism-based deep prediction network is used for receiving the fused feature vector, extracting time sequence features through a hybrid neural network and outputting a heart disease risk probability. The risk prediction and visualization output module is used for converting the heart disease risk probability into an understandable risk level and visually displaying the risk level. In the above technical solution, the individual differences of different populations and long-term health monitoring requirements are met.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and digital healthcare technology, and in particular to a training system for a cardiac condition risk prediction model. Background Technology

[0002] Cardiovascular disease is one of the leading causes of death worldwide. Early identification and risk assessment are crucial for preventing cardiac events such as myocardial infarction and heart failure. Traditional risk assessment tools, such as the Framingham Risk Score, are primarily based on static clinical variables (such as age, sex, blood pressure, and cholesterol levels), which have limited predictive accuracy and cannot reflect the dynamic changes in an individual's physiological state in real time.

[0003] In recent years, with the widespread adoption of wearable devices (such as smartwatches, ECG patches, and continuous glucose monitors), it has become possible to acquire continuous, multi-dimensional physiological signal data (such as electrocardiograms, photoplethysmography pulse waves, heart rate variability, respiratory rate, and blood oxygen saturation). These multimodal physiological signals contain rich information about the cardiovascular system's state, providing a data foundation for more accurate risk prediction.

[0004] Deep learning models, particularly convolutional neural networks and recurrent neural networks, have shown great potential in processing temporal physiological signals. However, current technologies face the following key challenges: Data heterogeneity and modal gaps: The frequency, quality, and modalities of signals acquired by different devices vary greatly (e.g., some devices only have PPG, while others have both ECG and PPG). In practical applications, users may not always wear all sensors, resulting in incomplete data modalities.

[0005] Individual differences and population generalization: A general model trained on large population data may have reduced predictive performance for specific individuals (such as people with special pathological features, different ages, genders, or lifestyles).

[0006] Conceptual drift and long-term adaptability: An individual's physiological state changes over time (e.g., disease progression, drug treatment, lifestyle changes), causing changes in data distribution (conceptual drift). Static models cannot adapt to these changes, and their predictive accuracy gradually decreases.

[0007] Privacy and computational efficiency: Medical data is highly sensitive. Centralizing everyone's data in the cloud for model retraining poses a risk of privacy breaches and incurs high computational and communication costs.

[0008] Therefore, there is an urgent need for a cardiac risk prediction model training system that can effectively integrate multimodal physiological signals and has continuous learning and personalized adaptation capabilities to overcome the shortcomings of the existing technologies. Summary of the Invention

[0009] This application provides a training system for a cardiac condition risk prediction model to meet the individual differences and long-term health monitoring needs of different population groups.

[0010] Firstly, a cardiac condition risk prediction model training system is provided, including: The data acquisition and preprocessing module is used to acquire multimodal physiological signal data and preprocess the data. The multimodal data adaptive fusion module is used to receive preprocessed multimodal physiological signal data, adaptively weight and fuse features of different modalities through a gated fusion mechanism, and output the fused feature vector. A deep prediction network based on an attention mechanism is used to receive the fused feature vector, extract temporal features through a hybrid neural network, and output the probability of heart disease risk. The risk prediction and visualization output module is used to convert the probability of heart disease risk into an understandable risk level and visualize it.

[0011] In the above technical solution, a data acquisition and preprocessing module is set up to collect multimodal physiological signal data and preprocess the data; a multimodal data adaptive fusion module is used to receive the preprocessed multimodal physiological signal data, adaptively weight and fuse the features of different modalities through a gating fusion mechanism, and output the fused feature vector; a deep prediction network based on an attention mechanism is used to receive the fused feature vector, extract temporal features through a hybrid neural network, and output the probability of heart disease risk; a risk prediction and visualization output module is used to convert the probability of heart disease risk into an understandable risk level and visualize it; thus meeting the needs of individual differences and long-term health monitoring for different populations.

[0012] In one specific implementation scheme, it also includes: Incremental learning and dynamic model update engine for continuous learning and personalized adaptation.

[0013] In one specific implementation scheme, the incremental learning and model dynamic update engine includes: The online sequence learning submodule is used to continuously fine-tune the model based on strategies such as elastic weight merging or online Bayesian updates. The Personalized Federated Learning submodule is used to collaboratively train the global model using the federated learning framework and generate user-personalized models based on model interpolation or meta-learning strategies. The Dynamic Model Selection and Integration submodule is used to dynamically select or weighted integrate the optimal model for prediction based on the similarity between the user's current data distribution and historical models.

[0014] In one specific implementation scheme, the gated fusion mechanism in the multimodal data adaptive fusion module specifically includes: For each time step, extract mode-specific feature vectors for each mode of data; Attention weights for each modality are calculated using an attention network, with the attention weights based on the information content and signal quality of the modality's feature vector. If data for a certain modality is missing, its attention weight is set to zero; The feature vectors of each modality are weighted and summed according to their attention weights to obtain the fused feature vector.

[0015] In one specific implementation scheme, the self-attention layer in the attention-based deep prediction network is a multi-head self-attention mechanism, which is used to capture the long-range dependencies between different time points in the physiological signal sequence and identify the time segments most critical for risk prediction.

[0016] In one specific implementation, the elastic weight merging method used in the online sequence learning submodule includes: Calculate the Fisher information matrix of the model parameters on the old data to measure the importance of each parameter to the old task; When updating model parameters using new data, a regularization term is added to the loss function to penalize deviations of important parameters from their old values.

[0017] In one specific implementation scheme, the model interpolation strategy in the personalized federated learning submodule is as follows: The final personalized model on the client is a weighted average of the global model and the locally fine-tuned model. The weighting coefficients are dynamically adjusted based on the amount of local data on the client or the difference between the local data and the global data distribution.

[0018] In one specific implementation scheme, the meta-learning strategy in the personalized federated learning submodule is: In federated learning, a model is trained to initialize parameters so that when a new user joins, only a small amount of local data is needed to perform a few steps of gradient descent on these initialized parameters to obtain a high-performance personalized model.

[0019] In one feasible implementation, a model repository is also included to store metadata for different versions of the base model, global model, and user models.

[0020] Secondly, a method for training a cardiac condition risk prediction model is provided, including the following steps: Multimodal physiological signal data were acquired using a data acquisition and preprocessing module, and the data were preprocessed. The preprocessed multimodal physiological signal data is received by the multimodal data adaptive fusion module, and the features of different modalities are adaptively weighted and fused through a gated fusion mechanism to output the fused feature vector. A deep prediction network based on an attention mechanism is used to receive the fused feature vector, extract temporal features through a hybrid neural network, and output the probability of heart disease risk. The risk prediction and visualization output module converts the probability of heart disease risk into an understandable risk level and displays it visually.

[0021] In the above technical solution, a data acquisition and preprocessing module is set up to collect multimodal physiological signal data and preprocess the data; a multimodal data adaptive fusion module is used to receive the preprocessed multimodal physiological signal data, adaptively weight and fuse the features of different modalities through a gating fusion mechanism, and output the fused feature vector; a deep prediction network based on an attention mechanism is used to receive the fused feature vector, extract temporal features through a hybrid neural network, and output the probability of heart disease risk; a risk prediction and visualization output module is used to convert the probability of heart disease risk into an understandable risk level and visualize it; thus meeting the needs of individual differences and long-term health monitoring for different populations. Attached Figure Description

[0022] Figure 1 This is a structural block diagram of the cardiac state risk prediction model training system provided in the embodiments of this application; Figure 2 A flowchart illustrating the training method for a cardiac state risk prediction model provided in this application embodiment. Detailed Implementation

[0023] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. Through these descriptions, the features and advantages of the present application will become clearer and more apparent.

[0024] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments. Although various aspects of embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless specifically indicated otherwise.

[0025] Furthermore, the technical features involved in the different embodiments of this application described below can be combined with each other as long as they do not conflict with each other.

[0026] To facilitate understanding of the cardiac state risk prediction model training system provided in this application embodiment, its application scenario is first explained. The cardiac state risk prediction model training system provided in this application embodiment is designed to meet the individual differences and long-term health monitoring needs of different population groups. Existing technologies face the following key challenges: Data heterogeneity and modal gaps: The frequency, quality, and modalities of signals collected by different devices vary greatly (e.g., some devices only have PPG, while others have both ECG and PPG). In practical applications, users may not always wear all sensors, resulting in incomplete data modalities. Individual differences and population generalization: A general model trained on large-scale population data may experience decreased predictive performance for specific individuals (e.g., people with specific pathological characteristics, different ages, genders, or lifestyles). Concept drift and long-term adaptability: An individual's physiological state changes over time (e.g., disease progression, drug treatment, lifestyle changes), leading to changes in data distribution (concept drift). Static models cannot adapt to these changes, and prediction accuracy gradually decreases. Privacy protection and computational efficiency: Medical data is highly sensitive. Centralizing all data in the cloud for model retraining poses a privacy risk and incurs high computational and communication costs. Therefore, there is an urgent need for a cardiac risk prediction model training system that can effectively integrate multimodal physiological signals and possess continuous learning and personalized adaptation capabilities to overcome the shortcomings of the existing technologies. To this end, this application provides a cardiac state risk prediction model training system to meet the individual differences and long-term health monitoring needs of different populations. The following detailed description, in conjunction with specific accompanying drawings, illustrates the system.

[0027] refer to Figure 1 and Figure 2 , Figure 1 This is a structural block diagram of the cardiac state risk prediction model training system provided in the embodiments of this application; Figure 2 A flowchart illustrating the training method for a cardiac state risk prediction model provided in this application embodiment.

[0028] exist Figure 1 In this application, an embodiment provides a cardiac condition risk prediction model training system, including: The data acquisition and preprocessing module is used to acquire multimodal physiological signal data and preprocess the data. The multimodal data adaptive fusion module is used to receive preprocessed multimodal physiological signal data, adaptively weight and fuse features of different modalities through a gated fusion mechanism, and output the fused feature vector. A deep prediction network based on an attention mechanism is used to receive the fused feature vector, extract temporal features through a hybrid neural network, and output the probability of heart disease risk. The risk prediction and visualization output module is used to convert the probability of heart disease risk into an understandable risk level and visualize it.

[0029] In the above technical solution, a data acquisition and preprocessing module is set up to collect multimodal physiological signal data and preprocess the data; a multimodal data adaptive fusion module is used to receive the preprocessed multimodal physiological signal data, adaptively weight and fuse the features of different modalities through a gating fusion mechanism, and output the fused feature vector; a deep prediction network based on an attention mechanism is used to receive the fused feature vector, extract temporal features through a hybrid neural network, and output the probability of heart disease risk; a risk prediction and visualization output module is used to convert the probability of heart disease risk into an understandable risk level and visualize it; thus meeting the needs of individual differences and long-term health monitoring for different populations.

[0030] Specifically, the beneficial effects include: Comprehensive and accurate data acquisition and processing lay a solid foundation: The data acquisition and preprocessing module can collect multimodal physiological signal data, covering various physiological information reflecting cardiac status. Compared with single-modal data, multimodal data provides a richer and more comprehensive perspective, enabling more accurate capture of subtle changes in cardiac status. Simultaneously, preprocessing the acquired data effectively removes noise, corrects erroneous data, and standardizes data formats, ensuring data quality and consistency. This provides a high-quality data foundation for subsequent analysis and modeling, significantly improving the accuracy and reliability of the entire system's predictions.

[0031] Adaptive fusion of multimodal data enhances feature representativeness: The adaptive multimodal data fusion module employs a gated fusion mechanism, enabling adaptive weighted fusion of features from different modalities. Different modalities of physiological signal data possess varying importance and characteristics in reflecting cardiac state. This mechanism dynamically adjusts the weights of each modality's features based on the data's inherent characteristics and importance, resulting in a fused feature vector that more comprehensively and accurately represents the heart's overall condition. This adaptive fusion approach avoids the limitations of traditional fixed-weight fusion methods, better adapting to individual physiological differences, improving feature representativeness and effectiveness, and providing more discriminative input for subsequent deep learning prediction.

[0032] Deep predictive networks combined with attention mechanisms enable accurate risk prediction: Deep predictive networks based on attention mechanisms extract temporal features through hybrid neural networks, deeply mining the potential information in multimodal fusion feature vectors. The attention mechanism automatically focuses on the most important features for predicting heart disease risk, making the model more attentive to key information and ignoring irrelevant interference, thereby improving the accuracy and specificity of predictions. Hybrid neural networks combine the advantages of various neural network structures, better handling complex temporal data and nonlinear relationships, further enhancing the model's ability to predict heart disease risk and providing more accurate output of heart disease risk probabilities, offering strong support for the early detection of potential heart disease risks.

[0033] Risk visualization enhances user understanding and engagement: The risk prediction and visualization module transforms the probability of heart disease risk into understandable risk levels and displays them visually. For non-professional users, the intuitive risk level display is easier to understand and accept than simple probability values, enabling users to quickly understand their own heart health status. Visual presentation methods, such as intuitive charts and color coding, present risk information in a concise and clear way, enhancing users' awareness and engagement with their own health, and encouraging them to take proactive health management measures, such as adjusting their lifestyle and seeking timely medical attention. This helps achieve early intervention and prevention of heart disease and improves overall health.

[0034] In one specific implementation scheme, it also includes: Incremental learning and dynamic model update engine for continuous learning and personalized adaptation.

[0035] In one specific implementation scheme, the incremental learning and model dynamic update engine includes: The online sequence learning submodule is used to continuously fine-tune the model based on strategies such as elastic weight merging or online Bayesian updates. The Personalized Federated Learning submodule is used to collaboratively train the global model using the federated learning framework and generate user-personalized models based on model interpolation or meta-learning strategies. The Dynamic Model Selection and Integration submodule is used to dynamically select or weighted integrate the optimal model for prediction based on the similarity between the user's current data distribution and historical models.

[0036] In one specific implementation scheme, the gated fusion mechanism in the multimodal data adaptive fusion module specifically includes: For each time step, extract mode-specific feature vectors for each mode of data; Attention weights for each modality are calculated using an attention network, with the attention weights based on the information content and signal quality of the modality's feature vector. If data for a certain modality is missing, its attention weight is set to zero; The feature vectors of each modality are weighted and summed according to their attention weights to obtain the fused feature vector.

[0037] In one specific implementation scheme, the self-attention layer in the attention-based deep prediction network is a multi-head self-attention mechanism, which is used to capture the long-range dependencies between different time points in the physiological signal sequence and identify the time segments most critical for risk prediction.

[0038] In one specific implementation, the elastic weight merging method used in the online sequence learning submodule includes: Calculate the Fisher information matrix of the model parameters on the old data to measure the importance of each parameter to the old task; When updating model parameters using new data, a regularization term is added to the loss function to penalize deviations of important parameters from their old values.

[0039] In one specific implementation scheme, the model interpolation strategy in the personalized federated learning submodule is as follows: The final personalized model on the client is a weighted average of the global model and the locally fine-tuned model. The weighting coefficients are dynamically adjusted based on the amount of local data on the client or the difference between the local data and the global data distribution.

[0040] In one specific implementation scheme, the meta-learning strategy in the personalized federated learning submodule is: In federated learning, a model is trained to initialize parameters so that when a new user joins, only a small amount of local data is needed to perform a few steps of gradient descent on these initialized parameters to obtain a high-performance personalized model.

[0041] In one feasible implementation, a model repository is also included to store metadata for different versions of the base model, global model, and user models.

[0042] Specifically, the beneficial effects include: In one specific implementation scheme, the cardiac state risk prediction model training system includes: Data acquisition and preprocessing module: responsible for acquiring raw physiological signals from various wearable or clinical devices and performing noise reduction, segmentation, alignment and standardization.

[0043] Multimodal data adaptive fusion module: A gated fusion mechanism is designed to adaptively weight and fuse physiological signal features of different modalities and handle the case of missing modalities.

[0044] Deep Prediction Network Based on Attention Mechanism: A hybrid neural network consisting of convolutional layers, recurrent neural network layers, and self-attention layers was constructed to extract temporal dependencies from the fused features and output the probability of heart disease risk.

[0045] Incremental learning and dynamic model update engine: This is the core innovation of this invention. The engine comprises three sub-modules: The online sequence learning submodule employs strategies based on Elastic Weight Consolidation (EWC) or online Bayesian methods to continuously fine-tune the model to adapt to new data.

[0046] Personalized Federated Learning Submodule: Under the premise of protecting user privacy, this module uses a federated learning framework to enable multiple clients (user devices) to collaboratively train a global model and supports the generation of personalized models.

[0047] Dynamic Model Selection and Integration Submodule: Based on the user's current data distribution and historical performance, dynamically selects or weights and integrates the optimal model from a set of base models or historical snapshots for prediction.

[0048] Risk prediction and visualization output module: Converts the probability values ​​output by the model into understandable risk levels and displays them to users or doctors through a graphical interface.

[0049] Furthermore, the core of the multimodal data adaptive fusion module is the gated fusion unit. For each time step, the feature vector of each modality is passed through an independent gated recurrent unit (GRU) or fully connected layer to generate a modality-specific context vector. Then, an attention network calculates the attention weight for each modality, which reflects the importance of that modality to the final prediction at the current time step. If data for a modality is missing, its attention weight is automatically reduced, and the system fuses only based on the available modalities. The fused feature vector is a weighted sum of the feature vectors of all available modalities.

[0050] Furthermore, the attention-based deep prediction network adopts a hierarchical structure: Low-level feature extractor: Uses a one-dimensional convolutional neural network (CNN) to extract local features from the raw or preprocessed signals of each modality.

[0051] Temporal dependency modeler: Uses bidirectional long short-term memory networks (Bi-LSTM) or GRUs to capture long-term temporal dependencies of physiological signals.

[0052] Self-attention layer: A multi-head self-attention mechanism is introduced on top of the output of Bi-LSTM, enabling the model to focus on the key time points most relevant to risk prediction.

[0053] Output layer: Finally, a fully connected layer and a sigmoid activation function are passed to output a risk probability value between 0 and 1.

[0054] Furthermore, the incremental learning and dynamic model update engine includes: 1. Online Sequence Learning Submodule: When new batches of data (e.g., user data from the past 24 hours) arrive, this submodule does not directly train the model from scratch with the new data, but instead uses methods based on Elastic Weights Combined (EWC) or online Bayesian updates.

[0055] The EWC method calculates the Fisher information matrix of important parameters on old data. When updating model parameters, it penalizes changes in parameters that are important to the old task (i.e., previously learned knowledge), thereby avoiding catastrophic forgetting while adapting to new data.

[0056] Online Bayesian methods treat model parameters as random variables and continuously update the posterior distribution of parameters using new data to achieve smooth model evolution.

[0057] 2. Personalized Federated Learning Submodule: The system employs a Federated Learning (FL) framework. Each user's device (client) trains the model locally using its own data, and only uploads model updates (such as gradients or weight differences) to the central server.

[0058] The central server aggregates model updates from multiple clients, generates an improved global model, and then distributes the global model to each client.

[0059] To accommodate individual differences, this invention introduces personalized strategies based on federated learning. Specifically, these include: Model interpolation: The final model on the client side is a weighted average of the global model and the locally fine-tuned model.

[0060] Meta-learning initialization: During federated learning, a model is trained to initialize its parameters to quickly adapt to new users. When a new user joins, these initialization parameters can be fine-tuned with only a small amount of data to obtain a personalized model.

[0061] 3. Dynamic Model Selection and Integration Submodule: The system periodically saves snapshots of the model at different points in time, forming a model library.

[0062] For each user's prediction request, the system calculates the similarity between the current user's data and the distribution of training data for each historical model (e.g., using the maximum mean difference, MMD).

[0063] The system can either select the model most similar to the current data distribution, or perform a weighted ensemble of predictions from multiple models, with the weights determined by similarity. This allows the system to dynamically switch to the model best suited to the user's current state.

[0064] In this embodiment, the beneficial effects include: High-precision prediction: Through adaptive fusion of multimodal data and deep networks based on attention mechanisms, complementary information in multi-source physiological signals can be fully extracted, significantly improving the accuracy of cardiac risk prediction.

[0065] Strong robustness: The gated fusion mechanism can effectively handle the problem of missing modes, ensuring that the system can still provide reliable predictions when some sensor data is unavailable.

[0066] Personalized adaptation: Through personalized federated learning and dynamic model selection, the system can generate customized prediction models for different users, overcoming the performance bottleneck of general models on individuals.

[0067] Long-term effectiveness: The incremental learning mechanism enables the model to continuously learn from new data, adapt to conceptual drift in individual physiological states, and ensure that predictive performance does not decline under long-term monitoring.

[0068] Privacy protection: The federated learning framework eliminates the need to centralize users' raw data for model training, effectively protecting users' privacy and security.

[0069] Computationally efficient: Online learning strategies reduce the need for storing and computing historical data, enabling model updates to be performed efficiently on resource-constrained edge devices such as smartphones.

[0070] In a specific feasible implementation, the system architecture mainly includes the following modules: Data Acquisition Layer: Users wear various wearable devices, such as smartwatches (collecting PPG and accelerometer data), ECG patches (collecting ECG data), and smart blood pressure monitors. These devices transmit raw data to the user's smartphone or gateway device via Bluetooth or Wi-Fi.

[0071] Edge computing layer (smartphone / gateway): Data preprocessing module: Denoises the received raw signal (e.g., using wavelet transform to remove baseline drift and power frequency interference), segments it (e.g., segments it into 5-minute segments), aligns it in time (ensuring that the timestamps of different modal data are consistent), and normalizes it.

[0072] Local model library: Stores user-customized models and / or lightweight global models.

[0073] Local incremental learner: Performs online sequence learning and fine-tunes the local model using the user's latest data.

[0074] Cloud service layer: Multimodal fusion and deep learning platform: responsible for more complex model training tasks, such as global model aggregation in federated learning.

[0075] Model repository: Stores metadata for different versions of base models, global models, and user models.

[0076] User data management: Store anonymized user metadata (such as age, gender, medical history) and model update information, rather than raw physiological signals.

[0077] 4. Application Layer: Provides a user interface and doctor portal to display risk prediction results, historical trends, and early warning information.

[0078] In one specific feasible implementation, the fusion process of the multimodal data adaptive fusion module includes: Modality-Specific Feature Extraction: For each temporal data segment (e.g., a 5-minute ECG signal), a separate feature extraction network is used. This network can consist of a one-dimensional CNN and a GRU. The CNN extracts local morphological features (such as R-wave peak value and waveform width), while the GRU captures temporal dynamics. The output is a modality-specific feature vector for each time step. , , .

[0079] Attention weight calculation: A shared attention network (usually a fully connected layer plus a softmax) receives feature vectors from all modalities and outputs the attention weights for each modality. , , The calculation of attention weights takes into account the information content of the features themselves as well as the reliability of the modality (such as the signal quality index).

[0080] Missing modality handling: If a modality is missing (e.g., ACC data is unavailable), its corresponding feature vector is set to an all-zero vector, and the attention network will automatically learn to reduce its weight.

[0081] Weighted fusion: The final fused feature vector The calculation is as follows: ; This fused feature vector captures the collaborative information of all available modes at time step t.

[0082] In a specific feasible implementation, the training and inference process of a deep prediction network is as follows: The specific structure of the network is as follows: Input layer: Receives the feature sequence after processing by the fusion module. ; Bi-LSTM layer: Uses multiple layers of bidirectional LSTM to further learn the contextual dependencies of the sequence. The hidden state of the last time step of the Bi-LSTM, or the mean / maximum of the hidden states of all time steps, can serve as a summary representation H of the entire sequence.

[0083] Multi-head self-attention layer: The output H of the Bi-LSTM is input into the multi-head self-attention mechanism. This mechanism allows the model to simultaneously focus on information from different positions in the sequence. Specifically, the query (Q), key (K), and value (V) matrices are obtained from H through a linear transformation. The attention score is calculated as follows: ; in It represents the dimension of the key vector. The outputs of multiple attention heads are concatenated and then passed through a linear layer.

[0084] Fully connected output layer: The output of the self-attention layer is flattened, passed through one or more fully connected layers, and finally output as a scalar using the Sigmoid activation function, which is the probability of heart disease risk P∈[0,1].

[0085] Training process: Loss function: Use the binary cross-entropy loss function because risk prediction is usually modeled as a binary classification problem (high risk vs. low risk).

[0086] ; in These are real labels (1 indicates a heart event or high risk, 0 indicates none). It is the probability predicted by the model.

[0087] Optimizer: Use the Adam optimizer to update parameters.

[0088] Regularization: Use Dropout and L2 regularization to prevent overfitting.

[0089] In a specific feasible implementation, the specific workflow of the incremental learning and dynamic model update engine includes: 1. Online sequence learning Assume the current model parameters are The newly arrived data batch is .

[0090] a. EWC method: In old data Compute each parameter on a subset diagonal elements of the Fisher information matrix It measures the importance of this parameter to older tasks.

[0091] The new loss function is defined as: ; in It's a loss based on new data. It is a hyperparameter used to control the strength of retaining old knowledge.

[0092] minimize Get the updated parameters .

[0093] b. Online Bayesian method: Treat the model parameters as a Gaussian distribution. .

[0094] When new data Upon arrival, update the posterior distribution using variational inference or Kalman filtering. The mean and variance of the parameters. When making predictions, the posterior mean is used as the parameter point estimate.

[0095] 2. Personalized Federated Learning Suppose there are K clients (users).

[0096] Local training: Each client k downloads the global model θg to its local machine. The model is trained for E epochs using the local data Dk, resulting in a local update θklocal.

[0097] Upload Update: The client will update locally. Alternatively, gradient information can be uploaded to the server. To protect privacy, updates can be encrypted using differential privacy encryption or homomorphic encryption.

[0098] Server aggregation: The server aggregates updates using the FedAvg algorithm. ; Alternatively, a weighted average can be calculated based on the amount of client data.

[0099] Personalization: Model Hybridization: The Client's Final Personalized Model for: ; Where β is the mixing coefficient, which controls the degree of personalization.

[0100] Meta-learning: In federated learning, the goal is to learn a model initialization. This allows for a high-performance personalized model to be obtained with only a few steps of gradient descent for a new user k. The loss function is designed as follows: ; That is, for each client, use its data pairs Perform N steps of gradient descent to obtain Then calculate The loss is calculated on the client-side test set, and the sum of the losses across all clients is minimized.

[0101] 3. Dynamic Model Selection and Integration Model library construction: The system periodically (e.g., monthly) saves snapshots of the current global model and / or user-customized models. .

[0102] Distribution similarity calculation: When it is necessary to make a prediction for user u, extract the user's data from the most recent period. .

[0103] calculate With each historical model Corresponding training data distribution similarity The maximum mean difference (MMD) or the log-likelihood based on the model output can be used as a similarity measure.

[0104] Model selection / integration: Selection method: Select the model with the highest similarity. Make predictions.

[0105] Ensemble method: Calculate each model weight ; Where τ is the temperature parameter. The final predicted probability is a weighted average of the predictions from each model.

[0106] In one specific implementation, to verify the effectiveness of the invention, experiments were conducted on a multimodal physiological signal dataset containing 10,000 subjects. The dataset included ECG, PPG, ACC signals, and labels indicating whether the subjects subsequently experienced a major adverse cardiac event (MACE).

[0107] The system proposed in this invention (denoted as Ours) is compared with the following baseline methods: Baseline 1: Framingham Risk Score based on Logistic Regression.

[0108] Baseline 2: Using a CNN-LSTM model with a single modality (ECG only).

[0109] Baseline 3: Uses a CNN-LSTM model with multimodal fusion, but without incremental learning and personalization capabilities (static model).

[0110] Evaluation metrics: Area under the curve (AUC), accuracy, precision, and recall.

[0111] Experimental results: Overall performance: Ours method achieved an AUC of 0.92, significantly higher than Baseline 1 (0.72), Baseline 2 (0.85), and Baseline 3 (0.88). This indicates that multimodal fusion and deep networks effectively improve prediction accuracy.

[0112] Personalization effect: Subjects were divided into multiple subgroups according to age, gender, and underlying diseases. Ours method maintained an AUC above 0.90 in all subgroups, while Baseline 3 showed greater AUC fluctuations (0.82-0.89) across different subgroups, indicating that the personalization mechanism of this invention effectively adapts to the differences among different populations.

[0113] Long-term adaptive simulation: A two-year data stream was simulated, in which the physiological characteristics of some subjects underwent conceptual drift after one year (such as worsening heart failure). Ours method, through incremental learning, maintained an AUC of 0.91 in the second year, while the AUC of Baseline 3 decreased to 0.84, demonstrating the advantage of this application in dealing with conceptual drift.

[0114] exist Figure 2 In this application, an embodiment provides a method for training a cardiac condition risk prediction model, including the following steps: Multimodal physiological signal data were acquired using a data acquisition and preprocessing module, and the data were preprocessed. The preprocessed multimodal physiological signal data is received by the multimodal data adaptive fusion module, and the features of different modalities are adaptively weighted and fused through a gated fusion mechanism to output the fused feature vector. A deep prediction network based on an attention mechanism is used to receive the fused feature vector, extract temporal features through a hybrid neural network, and output the probability of heart disease risk. The risk prediction and visualization output module converts the probability of heart disease risk into an understandable risk level and displays it visually.

[0115] In the above technical solution, a data acquisition and preprocessing module is set up to collect multimodal physiological signal data and preprocess the data; a multimodal data adaptive fusion module is used to receive the preprocessed multimodal physiological signal data, adaptively weight and fuse the features of different modalities through a gating fusion mechanism, and output the fused feature vector; a deep prediction network based on an attention mechanism is used to receive the fused feature vector, extract temporal features through a hybrid neural network, and output the probability of heart disease risk; a risk prediction and visualization output module is used to convert the probability of heart disease risk into an understandable risk level and visualize it; thus meeting the needs of individual differences and long-term health monitoring for different populations.

[0116] Those skilled in the art will know that this application can be implemented as a system, method, or computer program product.

[0117] Therefore, this disclosure can be implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this application can also be implemented as a computer program product in one or more computer-readable media, the computer-readable media containing computer-readable program code.

[0118] Any combination of one or more computer-readable media may be used. A computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this document, a computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device.

[0119] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application. Based on this, various substitutions and improvements can be made to this application, all of which fall within the protection scope of this application.

Claims

1. A training system for a cardiac condition risk prediction model, characterized in that, include: The data acquisition and preprocessing module is used to acquire multimodal physiological signal data and preprocess the data. The multimodal data adaptive fusion module is used to receive preprocessed multimodal physiological signal data, adaptively weight and fuse features of different modalities through a gated fusion mechanism, and output the fused feature vector. A deep prediction network based on an attention mechanism is used to receive the fused feature vector, extract temporal features through a hybrid neural network, and output the probability of heart disease risk. The risk prediction and visualization output module is used to convert the probability of heart disease risk into an understandable risk level and visualize it.

2. The cardiac condition risk prediction model training system according to claim 1, characterized in that, Also includes: Incremental learning and dynamic model update engine for continuous learning and personalized adaptation.

3. The cardiac condition risk prediction model training system according to claim 2, characterized in that, The incremental learning and dynamic model update engine includes: The online sequence learning submodule is used to continuously fine-tune the model based on strategies such as elastic weight merging or online Bayesian updates. The Personalized Federated Learning submodule is used to collaboratively train the global model using the federated learning framework and generate user-personalized models based on model interpolation or meta-learning strategies. The Dynamic Model Selection and Integration submodule is used to dynamically select or weighted integrate the optimal model for prediction based on the similarity between the user's current data distribution and historical models.

4. The cardiac condition risk prediction model training system according to claim 3, characterized in that, The gated fusion mechanism in the multimodal data adaptive fusion module specifically includes: For each time step, extract mode-specific feature vectors for each mode of data; Attention weights for each modality are calculated using an attention network, with the attention weights based on the information content and signal quality of the modality's feature vector. If data for a certain modality is missing, its attention weight is set to zero; The feature vectors of each modality are weighted and summed according to their attention weights to obtain the fused feature vector.

5. The cardiac condition risk prediction model training system according to claim 4, characterized in that, The self-attention layer in the deep prediction network based on the attention mechanism is a multi-head self-attention mechanism, which is used to capture the long-range dependencies between different time points in the physiological signal sequence and identify the most critical time segments for risk prediction.

6. The cardiac condition risk prediction model training system according to claim 5, characterized in that, The elastic weight merging method used in the online sequence learning submodule includes: Calculate the Fisher information matrix of the model parameters on the old data to measure the importance of each parameter to the old task; When updating model parameters using new data, a regularization term is added to the loss function to penalize deviations of important parameters from their old values.

7. The cardiac condition risk prediction model training system according to claim 6, characterized in that, The model interpolation strategy in the personalized federated learning submodule is as follows: The final personalized model on the client is a weighted average of the global model and the locally fine-tuned model. The weighting coefficients are dynamically adjusted based on the amount of local data on the client or the difference between the local data and the global data distribution.

8. The cardiac condition risk prediction model training system according to claim 7, characterized in that, The meta-learning strategy in the personalized federated learning submodule is as follows: In federated learning, a model is trained to initialize parameters so that when a new user joins, only a small amount of local data is needed to perform a few steps of gradient descent on these initialized parameters to obtain a high-performance personalized model.

9. The cardiac condition risk prediction model training system according to claim 8, characterized in that, It also includes a model repository, which stores metadata for different versions of base models, global models, and user models.

10. A method for training a cardiac condition risk prediction model, characterized in that, Includes the following steps: Multimodal physiological signal data were acquired using a data acquisition and preprocessing module, and the data were preprocessed. The preprocessed multimodal physiological signal data is received by the multimodal data adaptive fusion module, and the features of different modalities are adaptively weighted and fused through a gated fusion mechanism to output the fused feature vector. A deep prediction network based on an attention mechanism is used to receive the fused feature vector, extract temporal features through a hybrid neural network, and output the probability of heart disease risk. The risk prediction and visualization output module converts the probability of heart disease risk into an understandable risk level and displays it visually.

Images