Complex cardiovascular risk prediction model training method, prediction method, device and medium

Abstract

The application discloses a complex cardiovascular risk prediction model training method, a prediction method, equipment and a medium; the method comprises the following steps: acquiring training data containing sample multi-lead electrocardiogram signals, sample medical text data and sample disease labels; inputting the training data into an initial prediction model containing a morphological rhythm collaborative encoder and a semantic-driven feature aggregation module; outputting a prediction risk probability; calculating a target loss value based on the prediction risk probability and the sample disease label and updating the model parameters to obtain a trained complex cardiovascular risk prediction model; since the morphological rhythm collaborative encoder overcomes the scale difference between local morphology and global rhythm to alleviate feature fragmentation, and the semantic-driven feature aggregation module aggregates pathological features by taking medical text prior as an anchor point, the feature fragments are reconstructed, and the prediction accuracy of complex cardiovascular risk is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of medical data processing, and in particular to a training method, prediction method, device and medium for a complex cardiovascular risk prediction model. Background Technology

[0002] In today's public health system, cardiovascular disease has become a leading threat to human health due to its rapid onset, high mortality rate, and extremely short golden window for treatment, posing a severe challenge to the accuracy and timeliness of medical diagnostic systems. Faced with the increasing trend of population aging and the growing demand for health management, traditional manual interpretation is insufficient to meet the diagnostic needs of large-scale populations. The clinical and public health fields urgently need an efficient and automated method for automatic screening and early warning of such diseases. In recent years, deep learning models have made substantial progress in ECG-assisted diagnosis, reducing the cost of manual interpretation while improving diagnostic efficiency. Standard ECG diagnostic methods typically require simultaneous attention to local waveform details and global heart rhythm to identify cardiovascular syndromes caused by complex pathologies. Therefore, single-scale network architectures are often ineffective in ECG-assisted diagnostic scenarios.

[0003] Currently, mainstream parallel extraction paradigms for hybrid modeling are still limited by the severe feature fragmentation problem. Feature fragmentation essentially refers to the failure of multi-scale features to fully integrate and interact. This lack of interaction not only makes local morphology and global rhythm physically incompatible, but also causes complex pathological manifestations to be fragmented into discrete pieces due to the lack of a cohesive core. Specifically, at the physical level, the model lacks dynamic interaction between branches, making it difficult to achieve complementary calibration between morphology and rhythm. Mechanical feature splicing, such as channel-by-channel fusion, fails to effectively align the semantic spaces of features across different dimensions, leading to the easy neglect of crucial local details during fusion. Secondly, at the pathological level, this feature fragmentation exacerbates the fragmentation of pathological features. Simple lesions can maintain feature integrity under this architecture, but the complex features of compound complications are deconstructed into independent fragments, and the model lacks corresponding feature aggregation strategies, thus lacking the ability to classify complex syndromes. Summary of the Invention

[0004] The main objective of this invention is to provide a training method, prediction method, device, and medium for a complex cardiovascular risk prediction model, aiming to solve the technical problem that existing technologies struggle to overcome the feature fragmentation caused by scale differences, and are unable to accurately aggregate and reconstruct fragmented features in the diagnosis of complex complications.

[0005] To achieve the above objectives, the present invention provides a method for training a complex cardiovascular risk prediction model, the method comprising the following steps: Acquire training data, which includes sample multi-lead electrocardiogram signals, sample medical text data, and sample disease labels; An initial complex cardiovascular risk prediction model is constructed, and the training data is input into the initial complex cardiovascular risk prediction model to output the predicted cardiovascular risk probability. The complex cardiovascular risk prediction model includes a morphological rhythm co-encoder and a semantic-driven feature aggregation module. The target loss value is calculated based on the predicted cardiovascular risk probability and the sample disease label, and the parameters of the initial complex cardiovascular risk prediction model are updated based on the target loss value to obtain the trained complex cardiovascular risk prediction model.

[0006] Optionally, the initial complex cardiovascular risk prediction model is configured to perform the following steps: The morphological rhythm co-encoder is used to extract multi-scale features and perform cross-scale feature interaction on the multi-lead ECG signals of the sample to obtain panoramic visual features. The semantic-driven feature aggregation module performs attention feature aggregation on the sample medical text data and the panoramic visual features to obtain the target semantic aggregation features. The target semantic aggregate features are subjected to feature mapping processing to obtain the predicted cardiovascular risk probability.

[0007] Optionally, the morphological rhythm co-encoder includes a morphological feature encoding module, an adaptive gated spectrum injection module, a bidirectional cross-scale interaction module, and a multi-scale gated fusion module; the step of extracting multi-scale features and performing cross-scale feature interaction on the sample multi-lead ECG signals through the morphological rhythm co-encoder to obtain panoramic visual features includes: The multi-lead ECG signal of the sample is extracted by the multi-layer convolutional network of the morphological feature encoding module to obtain the first morphological feature, the second morphological feature and the third morphological feature. The third morphological feature is enhanced by the adaptive gated spectrum injection module to obtain deep rhythmic features. The bidirectional cross-scale interaction module performs cross-scale attention fusion based on the deep rhythm features and the second morphological features to generate semantic bridging features. The multi-scale gating fusion module stitches together the first morphological feature, the second morphological feature, and the semantic bridging feature to obtain multi-scale stitched features, and then performs channel-weighted aggregation on the multi-scale stitched features to obtain panoramic visual features.

[0008] Optionally, the feature enhancement processing of the third morphological features to obtain deep rhythmic features includes: The third morphological feature is subjected to Fourier transform processing to obtain complex spectral features; A frequency attention mask is generated based on the amplitude distribution of the complex spectral features, referring to the following formula: in, Represents a frequency attention mask. Represents the spectral characteristics of complex numbers. The amplitude distribution representing the characteristics of the complex spectrum. express Activation function, MLP stands for Multilayer Perceptron, It is a constant; Based on the frequency attention mask, the complex spectral features are enhanced, and the enhanced features are residually concatenated with the third morphological features to obtain deep rhythm features, as shown in the following formula: in, Indicates deep rhythm characteristics. Indicates the characteristics of the third form. This indicates the inverse Fourier transform operation. This indicates an element-wise multiplication operation.

[0009] Optionally, the step of performing cross-scale attention fusion based on the deep rhythm features and the second morphological features to generate semantic bridging features includes: The second morphological feature is aggregated by mean pooling and maximum pooling to generate a spatial saliency mask; The second morphological feature is weighted, filtered, and downsampled based on the spatial saliency mask to obtain the target morphological feature; A feature query vector is obtained by linearly mapping the deep rhythm features; The target morphological features are linearly mapped to obtain feature key vectors and feature value vectors; The matching degree is calculated based on the feature query vector and the feature key vector, and the feature value vector is weighted and aggregated based on the matching degree to obtain local morphological detail features, as shown in the following formula: in, To indicate local morphological details , and Let represent the linear projection matrices of the feature query vector, feature key vector, and eigenvalue vector, respectively. Represents the normalization function. Indicates the scaling factor. Indicates transpose. Indicates the morphological characteristics of the target; The deep rhythm features and the local morphological detail features are weighted and residually fused to obtain semantic bridging features, as shown in the following formula: in, This indicates semantic bridging features. Presentation layer normalization operation, This represents the scalar coefficients calculated through the gating network.

[0010] Optionally, the step of aggregating attention features of the sample medical text data and the panoramic visual features to obtain target semantic aggregation features includes: The sample medical text data is encoded using a pre-trained language model to obtain initial text embedding features; The initial text embedding features are mapped to obtain mapped text features; The mapped text features are concatenated with a preset category bias term to obtain a semantic query vector; Based on the semantic query vector, cross-attention calculation is performed on the panoramic visual features to obtain associated waveform features; The residuals of the associated waveform features and the semantic anchor query vector are added together and then subjected to layer normalization to obtain the target semantic aggregate features.

[0011] Optionally, the step of calculating a target loss value based on the predicted cardiovascular risk probability and the sample disease label, and updating the parameters of the initial complex cardiovascular risk prediction model based on the target loss value to obtain a trained complex cardiovascular risk prediction model includes: Calculate the prior probability of the risk category based on the sample quantity distribution in the disease labels of the samples; Based on the prior probability, the predicted cardiovascular risk probability is calibrated logarithmically to obtain the calibrated predicted risk probability. Extract the positive and negative samples corresponding to the disease labels of the samples, and assign asymmetric gradient adjustment weights to the positive and negative samples respectively; The target loss value is obtained by calculating the cross-entropy loss based on the gradient adjustment weights, the calibrated predicted risk probability, and the sample disease label; Gradient backpropagation is performed based on the target loss value to update the network parameters of the initial complex cardiovascular risk prediction model, thereby obtaining the trained complex cardiovascular risk prediction model.

[0012] Furthermore, to achieve the above objectives, this invention also proposes a method for predicting complex cardiovascular risks, the method comprising the following steps: Acquire the multi-lead ECG signals to be predicted and the medical text data to be predicted; The multi-lead electrocardiogram signal to be predicted and the medical text data to be predicted are input into a complex cardiovascular risk prediction model for risk prediction, and the complex cardiovascular risk prediction result is output. The complex cardiovascular risk prediction model is trained by the complex cardiovascular risk prediction model training method described above.

[0013] In addition, to achieve the above objectives, this application also proposes a complex cardiovascular risk prediction device, the device comprising: a memory, a processor, and a complex cardiovascular risk prediction program stored in the memory, the processor being used to run the complex cardiovascular risk prediction program, and the computer program being configured to implement the steps of the complex cardiovascular risk prediction model training method or the complex cardiovascular risk prediction method as described above.

[0014] In addition, to achieve the above objectives, this application also proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the complex cardiovascular risk prediction model training method or the complex cardiovascular risk prediction method described above.

[0015] This invention overcomes the limitations of traditional single-electrophysiological signal modeling by acquiring multimodal data including multi-lead ECG signals and medical text data, introducing cross-modal prior knowledge into complex pathological analysis. By constructing a morphological rhythm co-encoder in the prediction model, it achieves deep dynamic interaction between local subtle waveform morphology and global long-term heart rhythm at the physical level, effectively overcoming the drawbacks of feature fragmentation caused by scale differences. Furthermore, by constructing a semantic-driven feature aggregation module, it uses medical text prior data as stable semantic anchors to perform directional cross-search and aggregation calibration on highly discrete and fragmented complex complication features. Through the calculation of target loss values ​​and end-to-end forward inference in the prediction stage, the model accurately maps the probability of cardiovascular risk. Therefore, this invention fundamentally solves the technical problem of fragmented complex pathological representation, not only achieving accurate capture and reconstruction of rare lesions and complex complication features, significantly improving the prediction accuracy and interpretability of multi-label classification of complex cardiovascular syndromes, but also greatly reducing the discrimination ambiguity and misdiagnosis rate of artificial intelligence models when facing heterogeneous scale features and complex clinical scenarios. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the structure of a complex cardiovascular risk prediction device with a hardware operating environment involved in the embodiments of the present invention; Figure 2 This is a flowchart illustrating the first embodiment of the training method for a complex cardiovascular risk prediction model according to the present invention. Figure 3 This is a schematic diagram of the structure of the complex cardiovascular risk prediction model in the training method of the complex cardiovascular risk prediction model of the present invention; Figure 4 This is a flowchart illustrating the first embodiment of the complex cardiovascular risk prediction method of the present invention.

[0018] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0019] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0020] Reference Figure 1 , Figure 1 This is a schematic diagram of the complex cardiovascular risk prediction device structure in the hardware operating environment involved in the embodiments of the present invention.

[0021] like Figure 1As shown, the complex cardiovascular risk prediction device may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen and an input unit such as a keyboard; the user interface 1003 may also include standard wired and wireless interfaces. The network interface 1004 may optionally include standard wired and wireless interfaces (such as Wireless-Fidelity (Wi-Fi) interfaces). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk storage device. The memory 1005 may also optionally be a storage device independent of the aforementioned processor 1001.

[0022] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on complex cardiovascular risk prediction devices, which may include more or fewer components than illustrated, or combine certain components, or have different component arrangements.

[0023] like Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a network communication module, a user interface module, and a complex cardiovascular risk prediction program.

[0024] exist Figure 1 In the complex cardiovascular risk prediction device shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and the memory 1005 in the complex cardiovascular risk prediction device of the present invention can be set in the complex cardiovascular risk prediction device. The complex cardiovascular risk prediction device calls the complex cardiovascular risk prediction program stored in the memory 1005 through the processor 1001 and executes the complex cardiovascular risk prediction model training method or complex cardiovascular risk prediction method provided in the embodiments of the present invention.

[0025] This invention provides a method for training a complex cardiovascular risk prediction model, referring to... Figure 2 , Figure 2 This is a flowchart illustrating the first embodiment of the training method for the complex cardiovascular risk prediction model of the present invention.

[0026] In this embodiment, the training method for the complex cardiovascular risk prediction model includes the following steps: Step S10: Obtain training data.

[0027] It should be understood that the executing entity of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a computer, server, workstation, etc., or a terminal electronic device capable of realizing the above functions. The following uses a complex cardiovascular risk prediction device (hereinafter referred to as the prediction device) as an example to describe this embodiment and the following embodiments.

[0028] It should be noted that the training data refers to the historical multimodal dataset used to train the risk prediction model, including sample multi-lead electrocardiogram signals, sample medical text data, and sample disease labels.

[0029] The sample multi-lead ECG signal can be raw electrophysiological waveform data containing multiple lead channels (such as 12 leads) collected from the patient using equipment at a medical institution. The sample medical text data can be professional pathological description texts related to cardiovascular diseases, diagnostic guideline literature, or pre-built medical prior knowledge base texts. The sample disease label refers to the actual risk classification result corresponding to the sample multi-lead ECG signal, such as single-category or multi-category classification labels such as normal, myocardial infarction, and atrial fibrillation.

[0030] It should be understood that most current deep learning-based cardiovascular risk prediction methods employ parallel feature extraction. In this design, morphological and rhythmic features are isolated, lacking necessary dynamic interaction, and most feature fusion methods are simply hard-stitched together, ignoring the differences between different features. Under this mechanism, the model merely performs a simple superposition of feature dimensions, failing to achieve deep synergy between morphology and rhythm, resulting in a fragmented representation after fusion.

[0031] While constructing multi-resolution architectures has broadened the physical scope of feature extraction to some extent and significantly enhanced the model's ability to capture pathological information at different scales, these methods still have limitations when dealing with complex clinical scenarios. Existing multi-scale strategies, while enriching the physical receptive field, excessively pursue the independent capture of information at different frequencies, thus separating the detailed information of lesions from their overall contours. This feature fragmentation is further amplified, especially when dealing with complex lesions involving multiple lesions, where potential representations often scatter into highly discrete feature fragments. Due to the lack of stable aggregation modules, fragmented representations struggle to form a compact representation core, and simply stacking multiple scales in the model structure is insufficient to fundamentally reconstruct a complete pathological representation.

[0032] In summary, while existing hybrid neural network structures significantly improve the richness of ECG signal representation, they still fail to fundamentally solve the problem of feature fragmentation. This embodiment provides an effective solution to the aforementioned problem by constructing a bidirectional cross-scale interaction mechanism to achieve deep integration of morphology and rhythm at the physical level, and by utilizing a semantically driven feature aggregation module to reconstruct fragmented and complex pathological representations in the pathological space.

[0033] Step S20: Construct an initial complex cardiovascular risk prediction model, input the training data into the initial complex cardiovascular risk prediction model, and output the predicted cardiovascular risk probability.

[0034] It should be noted that the initial complex cardiovascular risk prediction model refers to a model whose parameters have not undergone back-optimization. The predicted cardiovascular risk probability refers to the numerical prediction probability score of the initial model for the input data belonging to each cardiovascular disease category.

[0035] A morphological rhythm co-encoder is a network module used to simultaneously extract and fuse local morphological waveform details and global long-term heart rhythm features from electrocardiogram (ECG) signals. A semantic-driven feature aggregation module is a network structure that uses textual semantic features as guiding anchors to perform cross-modal interactive retrieval and recombination aggregation of visual physiological features.

[0036] Understandably, the core challenge of multi-label ECG classification in predicting complex cardiovascular risks lies in overcoming the feature fragmentation caused by scale differences between local morphological details and global heart rhythm, and in aggregating fragmented features in the diagnosis of complex complications.

[0037] To address the aforementioned issues, this embodiment proposes a novel Multi-scale Spectral-Semantic Network (MSS_Net) to construct a complex cardiovascular risk prediction model.

[0038] It should be understood that by processing multi-scale physiological signals simultaneously through a morphological rhythm co-encoder, the problem of feature fragmentation caused by a single receptive field scale can be alleviated, and the effective fusion of local short-term waveforms and global long-term rhythms can be achieved. By introducing prior textual knowledge through a semantic-driven feature aggregation module, discrete physiological and pathological features can be accurately aggregated to corresponding semantic anchors, significantly enhancing the network's ability to identify and classify complex cardiovascular complications.

[0039] In some embodiments, the prediction device inputs the collected sample multi-lead ECG signals into a morphological rhythm co-encoder, which uses multi-layer serial or parallel convolution operators to extract local short-time waveform details of the signal, while combining time-frequency domain transformation operations or sequence attention networks to extract long-range global rhythm correlations of the signal.

[0040] The encoder internally constructs cross-scale interactive connections, weightedly concatenating and fusing local waveform features with global rhythm features to output a multi-scale panoramic physiological feature map. Sample medical text data is converted into text vectors by the text encoder and then input into the semantic-driven feature aggregation module.

[0041] The semantic-driven feature aggregation module uses text vectors as query conditions, performs cross-attention calculations on multi-scale panoramic physiological feature maps, and actively searches for lesion waveform regions that highly match the text descriptions. This process re-aggregates the originally fragmented physiological features into a complete and comprehensive pathological characterization. Finally, the classification task head outputs the predicted cardiovascular risk probability of the sample belonging to each disease category.

[0042] Step S30: Calculate the target loss value based on the predicted cardiovascular risk probability and the sample disease label, and update the parameters of the initial complex cardiovascular risk prediction model based on the target loss value to obtain the trained complex cardiovascular risk prediction model.

[0043] It should be noted that the target loss value refers to the numerical value used to quantify the error divergence between the predicted cardiovascular risk probability output by the model and the disease labels of the real samples. A well-trained complex cardiovascular risk prediction model refers to a deep learning model whose internal network parameters can accurately generalize and predict multimodal electrocardiogram data and reach a loss convergence state after multiple iterations of optimization.

[0044] In practical implementation, after the prediction device obtains the predicted cardiovascular risk probability from the network's forward output, it compares it with the actual disease labels of the samples. For example, algorithms such as cross-entropy loss function or asymmetric loss function with penalty weights can be used to calculate the magnitude of the difference between the two and obtain a specific target loss value.

[0045] Subsequently, an optimizer (such as the Adam optimizer or stochastic gradient descent optimizer) is used to calculate the gradient vector of each neuron node in the network based on the target loss value, and the weight parameters in the initial complex cardiovascular risk prediction model are updated through backpropagation algorithm according to the set learning rate.

[0046] The forward inference and backward update process is repeated continuously until the target loss value stabilizes or reaches the preset training period limit. At this point, the network parameters are fixed, resulting in a complex cardiovascular risk prediction model with accurate multi-label classification capabilities.

[0047] This embodiment uses a morphological rhythm co-encoder to extract and interact with signals at multiple scales, overcoming the scale difference between local microscopic waveforms and global macroscopic temporal features. A semantically driven feature aggregation module uses textual features as prior anchors to perform directional cross-searching and aggregation of discrete physiological features. Finally, an end-to-end network-supervised optimization mechanism is established through loss calculation and parameter iteration, thereby reconstructing complex pathological feature fragments at the pathological level. Overall, this effectively reduces the classification ambiguity and misdiagnosis rate of artificial intelligence systems when processing ECG signals of multiple complications, and improves the accuracy and discrimination ability of multi-label prediction results for cardiovascular risk.

[0048] Furthermore, based on the first embodiment described above, in the second embodiment of the complex cardiovascular risk prediction model training method of the present invention, the initial complex cardiovascular risk prediction model is configured to perform the following steps: Step S21: The morphological rhythm co-encoder is used to extract multi-scale features and perform cross-scale feature interaction on the sample multi-lead electrocardiogram signal to obtain panoramic visual features; Step S22: The semantic-driven feature aggregation module performs attention feature aggregation on the sample medical text data and the panoramic visual features to obtain the target semantic aggregation features; Step S23: Perform feature mapping processing on the target semantic aggregated features to obtain the predicted cardiovascular risk probability.

[0049] It should be noted that this embodiment formalizes multi-label electrocardiogram diagnosis as a process from the signal space. Mapping to the pathological semantic space .

[0050] Given input tensor The goal of this embodiment is to learn a parameterized function. To output the real labels Highly consistent prediction probability vectors. For example... Figure 3 As shown, Figure 3 This is a schematic diagram of the structure of a complex cardiovascular risk prediction model. In this embodiment, the complex cardiovascular risk prediction model MSS_Net adopts a hierarchical collaborative reasoning architecture, consisting of two core components: a morphological-rhythm collaborative encoder (MRCE) and a semantic-driven feature aggregation module.

[0051] Specifically, to address the feature fragmentation between local morphology and global rhythm, this embodiment constructs a bidirectional cross-scale interaction mechanism during the feature extraction stage: on the one hand, it uses AGSI to capture global rhythm, and on the other hand, it injects shallow details back into the deep semantic space, realizing cross-scale modeling of ECG signals from fine-grained waveforms to long-term rhythms. Secondly, to address the difficulty in diagnosing complex complications, this embodiment designs an SDFA module, which uses cross-modal interactive attention to calibrate potential features using pre-trained medical text priors as semantic anchors, thereby accurately classifying cardiovascular diseases in the pathological feature space and ensuring the semantic consistency and interpretability of the final diagnosis.

[0052] Furthermore, based on the second embodiment described above, in the third embodiment of the training method for the complex cardiovascular risk prediction model of the present invention, the morphological rhythm co-encoder includes a morphological feature encoding module, an adaptive gated spectrum injection module, a bidirectional cross-scale interaction module, and a multi-scale gated fusion module. Step S21 may include: Step S211: Extract features from the sample multi-lead ECG signal using the multi-layer convolutional network of the morphological feature encoding module to obtain the first morphological feature, the second morphological feature, and the third morphological feature. Step S212: Perform feature enhancement processing on the third morphological features through the adaptive gated spectrum injection module to obtain deep rhythm features; Step S213: The bidirectional cross-scale interaction module performs cross-scale attention fusion based on the deep rhythm features and the second morphological features to generate semantic bridging features; Step S214: The first morphological feature, the second morphological feature, and the semantic bridging feature are spliced ​​together by the multi-scale gating fusion module to obtain multi-scale spliced ​​features, and the multi-scale spliced ​​features are channel-weighted aggregated to obtain panoramic visual features.

[0053] It is understood that this embodiment solves the feature fragmentation problem by introducing Adaptive Gated Spectral Injection (AGSI) and Bidirectional Cross-scale Interaction (BCI); it designs a Semantically-Driven Feature Aggregation (SDFA) module, introduces a collaborative learning mechanism, and uses medical text priors as anchors to calibrate complex pathological features.

[0054] It should be noted that the complex cardiovascular risk prediction model consists of a morphological rhythm coencoder (MRCE) and a semantically driven feature aggregation (SDFA) module. The morphological rhythm coencoder comprises four components: a morphological feature encoding (MFE) for extracting local waveforms; an adaptive gated spectrum injection (AGSI) for capturing global frequency domain dependencies; a bidirectional cross-scale interactive coencoder (BCI) for bridging the semantic gap between deep and shallow layers; and a multi-scale feature aggregation module. The gated fusion module MGF is used for fusion. The SDFA module incorporates medical prior knowledge, uses textual semantic anchors to calibrate and classify visual features, and finally outputs multi-label prediction results.

[0055] To extract high-fidelity local waveform features from raw multi-lead ECG signals, this embodiment constructs a morphological feature encoding (MFE) module. Traditional convolutional networks are often limited by local receptive fields, causing subtle morphological changes in certain lesions to be ignored. To address this issue, this embodiment maximizes the richness and diversity of features through large-kernel grouped convolution and a channel competition mechanism.

[0056] Considering that the 12-lead ECG signal records independent electrophysiological activities of the heart from different spatial angles, this embodiment uses grouped convolution to independently model each lead to avoid destroying the spatial structural integrity between leads during the shallow feature extraction stage. The MFE module consists of a 3-layer convolutional network, and the output features of each layer are denoted as the first morphological features. Second morphological characteristics and third-form characteristics .

[0057] This embodiment uses a large 31×1 convolution kernel. This large receptive field design can cover the complete heartbeat morphological cycle, thereby accurately capturing long-duration electrocardiogram waveforms (such as QRS waves).

[0058] Furthermore, to avoid excessive focus on common pathological patterns by numerous channels while neglecting rare lesions, this embodiment integrates a GRN module to stimulate competition between channels. By adaptively suppressing redundant homogeneous feature responses, the network is incentivized to capture more diverse fine-grained morphological details, preserving high-fidelity pathological features for accurate identification of rare lesions in subsequent stages. Specifically, given the original input ECG signal... After downsampling and encoding operations, the final output morphological features This can be formally expressed as Formula 1 below: in, Indicates the second morphological characteristic; This represents a global response normalization operation, used to stimulate competition between channels and adaptively suppress redundant homogeneous feature responses; Indicates the activation function; This represents the intermediate features of the original input electrocardiogram signal; This represents the original input electrocardiogram signal; Indicates a downsampling operation; This represents a large kernel depthwise convolution; For layer normalization; and These represent the 1×1 pointwise convolutional matrices responsible for channel expansion and projection in the inverted bottleneck layer, respectively; This represents the residual connection path.

[0059] Furthermore, based on the third embodiment described above, in the fourth embodiment of the complex cardiovascular risk prediction model training method of the present invention, step S212 may include: Step S2121: Perform Fourier transform processing on the third morphological feature to obtain complex spectral features; Step S2122: Generate a frequency attention mask based on the amplitude distribution of the complex spectral features; Step S2123: Enhance the complex spectral features based on the frequency attention mask, and perform residual connection between the enhanced features and the third morphological features to obtain deep rhythm features.

[0060] It should be noted that, in order to overcome the limitations of pure convolutional structures in capturing long-range heart rhythm features, this embodiment designs an adaptive gated spectrum injection module. Its core function is to use the macroscopic perception capability of the spectrum space to explicitly extract global rhythm features with diagnostic value and inject them back into the deep feature space as contextual information, thereby achieving deep coordination between local waveform morphology and global heart rhythm.

[0061] Specifically, this embodiment will The data is fed into the third convolutional layer to obtain deep features. First of all, The complex spectrum is obtained by performing a Fourier transform. This is used to store all information about the electrocardiogram (ECG) signal in the frequency domain. Considering that not all frequency bands contribute to diagnosis, this embodiment constructs a lightweight spectrum-gated network to achieve adaptive filtering. This network generates a frequency attention mask based on the amplitude distribution of the spectrum. Dynamically enhance pathology-related rhythmic frequency bands while suppressing background noise, as shown in Formula 2 below: in, Indicates frequency attention mask; Represents the spectral characteristics of complex numbers; Then it means Activation function, MLP stands for Multilayer Perceptron, It is a minimal constant used for numerical stability.

[0062] Finally, the enhanced features are restored back to the time domain through inverse Fourier transform, and then injected into the original features through residual connections to generate deep rhythmic features containing global rhythms. Refer to Formula 3 below: in, Indicates deep rhythm characteristics; Indicates the third form of characteristics; This indicates the inverse Fourier transform operation; Represents the spectral characteristics of complex numbers; This indicates an element-wise multiplication operation.

[0063] Through the aforementioned spectral injection mechanism, the model successfully integrates abstract deep semantics with macroscopic heartbeat rhythms. This not only compensates for the limitations of convolutional networks in their receptive field of view but also ensures the consistency of feature representation in terms of morphological integrity and rhythmic coherence. This provides deep rhythmic features containing global temporal dependencies for subsequent cross-scale interactions. .

[0064] Furthermore, based on the third embodiment described above, in the fifth embodiment of the complex cardiovascular risk prediction model training method of the present invention, step S213 may include: Step S2131: Perform mean pooling and max pooling on the second morphological feature to generate a spatial saliency mask; Step S2132: Perform weighted filtering and downsampling on the second morphological feature based on the spatial saliency mask to obtain the target morphological feature; Step S2133: Perform linear mapping on the deep rhythm features to obtain the feature query vector; Step S2134: Perform linear mapping on the target morphological features to obtain feature key vector and feature value vector; Step S2135: Calculate the matching degree based on the feature query vector and the feature key vector, and perform weighted aggregation on the feature value vector based on the matching degree to obtain local morphological detail features; Step S2136: Perform weighted residual fusion on the deep rhythm features and the local morphological detail features to obtain semantic bridging features.

[0065] It should be noted that the encoder effectively extracts multi-level features through its hierarchical structure, but deep features... shallow features Significant semantic gaps still exist, with deep features often losing morphological details that are crucial for small lesions.

[0066] To overcome this deficiency, this embodiment proposes a BCI module. Its core idea is to use deep rhythmic features as active query signals to actively retrieve and recover relevant morphological detail features from shallow space. This information is then precisely injected into deep features, and subsequently fused to generate semantic bridging features that combine global rhythm and local morphology. .

[0067] Specifically, the interaction process includes three stages: gating filtering, cross-scale retrieval, and semantic reconstruction.

[0068] First, to avoid background noise in shallow features interfering with the interaction process, this embodiment introduces a spatial saliency gating mechanism. A spatial saliency mask is generated by aggregating the average and maximum values ​​of the features. For the original shallow features Weighted filtering and downsampling are performed to obtain denoised morphological key pairs. .

[0069] Secondly, this embodiment designs a cross-scale attention mechanism to generate local waveform detail features. This embodiment will utilize deep rhythmic characteristics. As a query vector, the denoised shallow morphological features As key and value vectors, the model can dynamically identify and aggregate local waveform details most relevant to the current ECG rhythm state by calculating the matching degree between them. Refer to Formula 4 below: in, To indicate local morphological details , , Let represent the linear projection matrices of the feature query vector, feature key vector, and eigenvalue vector, respectively, and Softmax represent the normalization function. This represents the scaling factor, and T represents the transpose. It indicates the morphological characteristics of the target.

[0070] Finally, considering and Because these two features reside in different semantic spaces, simple linear superposition may prevent them from being deeply integrated. Therefore, this embodiment introduces an adaptive residual injection mechanism to dynamically control the inflow of morphological information based on diagnostic needs. Specifically, this embodiment first concatenates the two features and then uses a gating network to calculate scalar coefficients. This is used to quantitatively evaluate the value of current morphological details to deep semantics. Subsequently, this coefficient is used to weight the morphological details and injected into the backbone network as residuals to generate the final semantic bridging features. Refer to Formula 5 below: in, This indicates semantic bridging features. Presentation layer normalization operation, This represents the scalar coefficients calculated through the gating network.

[0071] In one embodiment, information at all granularities must be integrated to generate the final diagnostic characterization. In the multi-scale gated fusion module, the highest-resolution feature F1 retained from the first stage is reintroduced to ensure that minute lesions are not missed.

[0072] Specifically, this embodiment uses a multi-scale feature set. Align to a uniform dimension and stitch together to obtain To adaptively adjust the contribution rate of features at different scales, this embodiment introduces a context-gated perceptual network, compresses spatial information through global average pooling, and utilizes an MLP to generate dynamic channel weights. .

[0073] Final visual features The result is obtained through channel-weighted aggregation, as shown in Formula 6 below: This mechanism ensures that the model can dynamically focus on the most diagnostically valuable feature level based on the characteristics of the input signal, thereby providing a high-fidelity panoramic pathological representation for subsequent decision-making.

[0074] Furthermore, based on the second embodiment described above, in the sixth embodiment of the complex cardiovascular risk prediction model training method of the present invention, step S22 may include: Step S221: Encode the sample medical text data using a pre-trained language model to obtain initial text embedding features; Step S222: Map the initial text embedding features to obtain mapped text features; Step S223: Concatenate the mapped text features with a preset category bias term to obtain a semantic query vector; Step S224: Perform cross-attention calculation on the panoramic visual features based on the semantic query vector to obtain associated waveform features; Step S225: Add the residuals of the associated waveform features and the semantic anchor query vector, and perform layer normalization to obtain the target semantic aggregation features.

[0075] It should be noted that although the Morphological Rhythm Co-encoder (MRCE) alleviates feature fragmentation at the ECG signal level, when faced with complex complications, pathological features are often deconstructed into highly discrete independent fragments, which can easily lead to ambiguity in the classifier's decision-making. To address this challenge, this embodiment introduces semantic-driven feature aggregation, using prior medical knowledge as an anchor point for semantic cohesion to perform targeted aggregation and reconstruction of fragmented features.

[0076] Specifically, the process consists of two steps: semantic anchor construction and cross-modal alignment. First, this embodiment utilizes a pre-trained language model to extract... Text embeddings of each category And freeze during training. To preserve the inherent semantic topology among medical terms, this embodiment introduces a learnable projection module to generate the final semantic query vector in order to align semantic features to the ECG signal space. Refer to Formula 7 below: in, This indicates mapping the text dimension to the visual dimension. , It is a learnable class bias term used to fine-tune the feature distribution of a specific class, thereby achieving accurate capture of features with slight individual differences.

[0077] Subsequently, this embodiment will generate semantic anchors. As a semantic query vector, the fused panoramic visual features As semantic key vectors and semantic value vectors, the model actively searches for and aggregates relevant pathological waveform features based on semantic prototypes through a multi-head cross-attention mechanism, as shown in Formula 8 below: in, Indicates associated waveform characteristics, This represents the semantic query vector that serves as the semantic anchor. , , It is a linear projection matrix.

[0078] To prevent gradient vanishing and promote training stability, this embodiment employs residual connections and layer normalization after the attention layer, and outputs the predicted probability through the classification head. Refer to Formula 9 below: (9) in, Represents the target semantic aggregation features. This represents the activation function. Indicates the predicted probability. The weight matrix represents the classification head. Indicates category bias.

[0079] This mechanism enables each text anchor representing a specific disease to actively capture the most relevant waveform features across the entire image through explicit interaction between pathological text descriptions and electrocardiogram features. This alleviates the feature fragmentation problem in complex complications and ensures that the final diagnostic results are both accurate and medically rigorous.

[0080] Furthermore, based on the above embodiments, in the seventh embodiment of the complex cardiovascular risk prediction model training method of the present invention, step S30 may include: Step S301: Calculate the prior probability of the risk category based on the sample quantity distribution in the disease label of the sample; Step S302: Perform probability logarithmic calibration calculation on the predicted cardiovascular risk probability based on the prior probability to obtain the calibrated predicted risk probability; Step S303: Extract the positive and negative samples corresponding to the disease labels of the samples, and assign asymmetric gradient adjustment weights to the positive and negative samples respectively; Step S304: Calculate the cross-entropy loss based on the gradient adjustment weights, the calibrated predicted risk probability, and the sample disease label to obtain the target loss value; Step S305: Perform gradient backpropagation based on the target loss value to update the network parameters of the initial complex cardiovascular risk prediction model, and obtain the trained complex cardiovascular risk prediction model.

[0081] It should be noted that, in response to the problem of fragmented features in complex syndromes, this embodiment optimizes the model from two dimensions: decision boundary calibration and gradient optimization.

[0082] First, to counteract the systematic bias caused by class imbalance, this embodiment introduces prior-aware calibration at the logit layer. This is achieved by injecting class prior probabilities. This forces an increase in decision-making weight for scarce lesions. Let... Let be the number of samples in the k-th category. Let be the total number of samples, then the prior probability of the class is... = / .remember The original logit value is the model output, and the calibrated logit value is... Defined as: (10) in, This represents the calibrated logit value; This represents the raw logit value output by the model; It is a hyperparameter for adjusting intensity; Let represent the prior probability of the k-th risk category.

[0083] Subsequently, to prevent a large number of simple negative samples from dominating the gradient direction, this embodiment employs asymmetric loss for optimization. This allows the model to use two completely different weighting strategies for positive and negative samples, and to severely suppress the gradient of easily classified background classes. Specifically, based on the calibrated probability... for The total loss is defined as: (11) in, Indicates the target loss value; It is the smoothed label of the k-th risk category. Indicates the probability after calibration; The weight adjustment parameter represents the weight of the positive samples; This represents the weight adjustment parameter for negative samples.

[0084] For example, setting For 2 and It is 0. This means that for positive samples ( For negative samples (where the value is 0), this embodiment preserves the complete gradient flow to ensure recall; while for negative samples (where the value is 0), the gradient flow is preserved to ensure recall. (For example, if the weight is 2), the model will automatically reduce the weight of simple samples and focus on mining difficult negative samples.

[0085] refer to Figure 4 , Figure 4 This is a flowchart illustrating the first embodiment of the complex cardiovascular risk prediction method of the present invention.

[0086] In this embodiment, the complex cardiovascular risk prediction method includes the following steps: Step S100: Obtain the multi-lead ECG signal to be predicted and the medical text data to be predicted.

[0087] It should be understood that the executing entity of this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a computer, server, workstation, etc., or a terminal electronic device capable of realizing the above functions. The following uses a complex cardiovascular risk prediction device (hereinafter referred to as the prediction device) as an example to describe this embodiment and the following embodiments.

[0088] It should be noted that the multi-lead ECG signal to be predicted refers to the real-time or offline multi-channel ECG waveform data of the subject that needs to be analyzed for disease diagnosis. The medical text data to be predicted can be electronic medical record description text related to the subject, chief symptom records, or standardized text from a medical prior knowledge base.

[0089] In practical implementation, the prediction device can receive real-time electrocardiogram (ECG) data from the subject collected by an electrocardiograph or wearable monitoring device via a wired or wireless communication interface, and perform standardization processing such as baseline filtering and resampling to obtain the multi-lead ECG signal to be predicted. Simultaneously, the device retrieves the subject's electronic medical record from the hospital information system, or extracts a pre-structured cardiovascular disease text knowledge base, and forms the medical text data to be predicted after word segmentation and vectorization.

[0090] Step S200: Input the multi-lead ECG signal to be predicted and the medical text data to be predicted into the complex cardiovascular risk prediction model for risk prediction, and output the complex cardiovascular risk prediction result. The complex cardiovascular risk prediction model is trained by the complex cardiovascular risk prediction model training method described above.

[0091] In practice, the prediction device pre-trains an initial complex cardiovascular risk prediction model. The training method for the complex cardiovascular risk prediction model is basically the same as the various embodiments of the training method for the complex cardiovascular risk prediction model described above, and will not be repeated here.

[0092] In some embodiments, the prediction device loads the weight parameters of a pre-trained complex cardiovascular risk prediction model into local memory or a cloud inference server. Subsequently, the acquired multi-lead ECG signals and medical text data to be predicted are synchronously input into the model for forward computation (without backward gradient propagation). The model extracts multi-scale features from the ECG signals and combines them with the medical text for cross-attention aggregation. After passing through a final classification layer, it outputs a complex cardiovascular risk prediction result containing probability distributions for multiple disease categories.

[0093] This embodiment uses a morphological rhythm co-encoder to extract and interact with signals at multiple scales, overcoming the scale difference between local microscopic waveforms and global macroscopic temporal features. A semantically driven feature aggregation module uses textual features as prior anchors to perform directional cross-searching and aggregation of discrete physiological features. Finally, an end-to-end network-supervised optimization mechanism is established through loss calculation and parameter iteration, thereby reconstructing complex pathological feature fragments at the pathological level. Overall, this effectively reduces the classification ambiguity and misdiagnosis rate of artificial intelligence systems when processing ECG signals of multiple complications, and improves the accuracy and discrimination ability of multi-label prediction results for cardiovascular risk.

[0094] Furthermore, this invention also proposes a computer-readable storage medium storing a complex cardiovascular risk prediction program, which, when executed by a processor, implements the steps of the complex cardiovascular risk prediction model training method or the complex cardiovascular risk prediction method described above.

[0095] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, 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 devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0096] The aforementioned computer-readable storage medium may be included in a complex cardiovascular risk prediction device; or it may exist independently and not be assembled into a complex cardiovascular risk prediction device.

[0097] Furthermore, this invention also proposes a computer program product, including a complex cardiovascular risk prediction program, which, when executed by a processor, implements the steps of the complex cardiovascular risk prediction model training method or the complex cardiovascular risk prediction method as described above.

[0098] The specific implementation of the computer program product of the present invention is basically the same as the above-mentioned training method for complex cardiovascular risk prediction model or various embodiments of complex cardiovascular risk prediction method, and will not be repeated here.

[0099] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0100] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0101] In addition, for technical details not described in detail in this embodiment, please refer to the training method for complex cardiovascular risk prediction models or the complex cardiovascular risk prediction method provided in any embodiment of the present invention, which will not be repeated here.

[0102] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0103] It should be noted that the user information (including but not limited to user device information, user personal information, user location information, user behavior information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0104] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0105] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0106] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A training method for a complex cardiovascular risk prediction model, characterized in that, The training method for the complex cardiovascular risk prediction model includes: Acquire training data, which includes sample multi-lead electrocardiogram signals, sample medical text data, and sample disease labels; An initial complex cardiovascular risk prediction model is constructed, and the training data is input into the initial complex cardiovascular risk prediction model to output the predicted cardiovascular risk probability. The complex cardiovascular risk prediction model includes a morphological rhythm co-encoder and a semantic-driven feature aggregation module. The target loss value is calculated based on the predicted cardiovascular risk probability and the sample disease label, and the parameters of the initial complex cardiovascular risk prediction model are updated based on the target loss value to obtain the trained complex cardiovascular risk prediction model.

2. The training method for a complex cardiovascular risk prediction model as described in claim 1, characterized in that, The initial complex cardiovascular risk prediction model is configured to perform the following steps: The morphological rhythm co-encoder is used to extract multi-scale features and perform cross-scale feature interaction on the multi-lead ECG signals of the sample to obtain panoramic visual features. The semantic-driven feature aggregation module performs attention feature aggregation on the sample medical text data and the panoramic visual features to obtain the target semantic aggregation features. The target semantic aggregate features are subjected to feature mapping processing to obtain the predicted cardiovascular risk probability.

3. The training method for a complex cardiovascular risk prediction model as described in claim 2, characterized in that, The morphological rhythm co-encoder includes a morphological feature encoding module, an adaptive gated spectrum injection module, a bidirectional cross-scale interaction module, and a multi-scale gated fusion module; the morphological rhythm co-encoder performs multi-scale feature extraction and cross-scale feature interaction on the sample multi-lead ECG signals to obtain panoramic visual features, including: The multi-lead ECG signal of the sample is extracted by the multi-layer convolutional network of the morphological feature encoding module to obtain the first morphological feature, the second morphological feature and the third morphological feature. The third morphological feature is enhanced by the adaptive gated spectrum injection module to obtain deep rhythmic features. The bidirectional cross-scale interaction module performs cross-scale attention fusion based on the deep rhythm features and the second morphological features to generate semantic bridging features. The multi-scale gating fusion module stitches together the first morphological feature, the second morphological feature, and the semantic bridging feature to obtain multi-scale stitched features, and then performs channel-weighted aggregation on the multi-scale stitched features to obtain panoramic visual features.

4. The training method for a complex cardiovascular risk prediction model as described in claim 3, characterized in that, The feature enhancement processing of the third morphological features to obtain deep rhythmic features includes: The third morphological feature is subjected to Fourier transform processing to obtain complex spectral features; A frequency attention mask is generated based on the amplitude distribution of the complex spectral features, referring to the following formula: in, This represents a frequency attention mask. Represents the spectral characteristics of complex numbers. The amplitude distribution representing the characteristics of the complex spectrum. express Activation function, MLP stands for Multilayer Perceptron, It is a constant; Based on the frequency attention mask, the complex spectral features are enhanced, and the enhanced features are residually concatenated with the third morphological features to obtain deep rhythm features, as shown in the following formula: in, Indicates deep rhythm characteristics. Indicates the characteristics of the third form. This indicates the inverse Fourier transform operation. This indicates an element-wise multiplication operation.

5. The training method for a complex cardiovascular risk prediction model as described in claim 3, characterized in that, The step of performing cross-scale attention fusion based on the deep rhythm features and the second morphological features to generate semantic bridging features includes: The second morphological feature is aggregated by mean pooling and maximum pooling to generate a spatial saliency mask; The second morphological feature is weighted, filtered, and downsampled based on the spatial saliency mask to obtain the target morphological feature; A feature query vector is obtained by linearly mapping the deep rhythm features; The target morphological features are linearly mapped to obtain feature key vectors and feature value vectors; The matching degree is calculated based on the feature query vector and the feature key vector, and the feature value vector is weighted and aggregated based on the matching degree to obtain local morphological detail features, as shown in the following formula: in, To indicate local morphological details , and Let represent the linear projection matrices of the feature query vector, feature key vector, and eigenvalue vector, respectively. Represents the normalization function. Indicates the scaling factor. Indicates transpose. Indicates the morphological characteristics of the target; The deep rhythm features and the local morphological detail features are weighted and residually fused to obtain semantic bridging features, as shown in the following formula: in, This represents semantic bridging features. Presentation layer normalization operation, This represents the scalar coefficients calculated through the gating network.

6. The training method for a complex cardiovascular risk prediction model as described in claim 2, characterized in that, The attention feature aggregation of the sample medical text data and the panoramic visual features to obtain target semantic aggregation features includes: The sample medical text data is encoded using a pre-trained language model to obtain initial text embedding features; The initial text embedding features are mapped to obtain mapped text features; The mapped text features are concatenated with a preset category bias term to obtain a semantic query vector; Based on the semantic query vector, cross-attention calculation is performed on the panoramic visual features to obtain associated waveform features; The residuals of the associated waveform features and the semantic anchor query vector are added together and then subjected to layer normalization to obtain the target semantic aggregate features.

7. The training method for a complex cardiovascular risk prediction model as described in claim 1, characterized in that, The process involves calculating a target loss value based on the predicted cardiovascular risk probability and the sample disease label, and updating the parameters of the initial complex cardiovascular risk prediction model based on the target loss value to obtain a trained complex cardiovascular risk prediction model, including: Calculate the prior probability of the risk category based on the sample quantity distribution in the disease labels of the samples; Based on the prior probability, the predicted cardiovascular risk probability is calibrated logarithmically to obtain the calibrated predicted risk probability. Extract the positive and negative samples corresponding to the disease labels of the samples, and assign asymmetric gradient adjustment weights to the positive and negative samples respectively; The target loss value is obtained by calculating the cross-entropy loss based on the gradient adjustment weights, the calibrated predicted risk probability, and the sample disease label; Gradient backpropagation is performed based on the target loss value to update the network parameters of the initial complex cardiovascular risk prediction model, thereby obtaining the trained complex cardiovascular risk prediction model.

8. A method for predicting complex cardiovascular risk, characterized in that, The complex cardiovascular risk prediction method includes: Acquire the multi-lead ECG signals to be predicted and the medical text data to be predicted; The multi-lead electrocardiogram signal to be predicted and the medical text data to be predicted are input into a complex cardiovascular risk prediction model for risk prediction, and the complex cardiovascular risk prediction result is output. The complex cardiovascular risk prediction model is obtained by training the complex cardiovascular risk prediction model training method as described in any one of claims 1 to 7.

9. A complex cardiovascular risk prediction device, characterized in that, The complex cardiovascular risk prediction device includes: a memory, a processor, and a complex cardiovascular risk prediction program stored in the memory. The processor is used to run the complex cardiovascular risk prediction program, which is configured to implement the complex cardiovascular risk prediction model training method as described in any one of claims 1 to 7 or the complex cardiovascular risk prediction method as described in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a complex cardiovascular risk prediction program, which, when executed by a processor, implements the complex cardiovascular risk prediction model training method as described in any one of claims 1 to 7 or the complex cardiovascular risk prediction method as described in claim 8.

Images