CN-122025093-A - Cardiovascular disease prediction method and device based on electrocardiographic large model, electronic equipment and medium

CN122025093ACN 122025093 ACN122025093 ACN 122025093ACN-122025093-A

Abstract

The disclosure relates to the technical field of medical signal processing and artificial intelligence, in particular to a cardiovascular disease prediction method and device based on an electrocardiographic large model, electronic equipment and medium. The method comprises the steps of inputting a pre-trained electrocardio segment set into an electrocardio large model for self-supervision pre-training to obtain the pre-trained electrocardio large model, pre-training by utilizing a potential representation feature sequence obtained by extracting features of multi-lead electrocardio segments, and minimizing a multi-task loss function, wherein the multi-task loss function comprises context contrast loss and mask language modeling loss, inputting a fine-tuning electrocardio segment set into the pre-trained electrocardio large model, and performing model fine-tuning on the pre-trained electrocardio large model according to a disease prediction result and a real disease label to obtain a cardiovascular disease prediction model so as to realize cardiovascular disease prediction of an individual to be detected. According to the method, the analysis object trained by the model is migrated from the original electrocardiosignal to the potential representation feature sequence, the model does not need to consume a large amount of resources to construct irrelevant details and noise, and the robustness is good.

Inventors

ZHAO HUIYING
YANG YUEDONG
LIN SIYING

Assignees

中山大学孙逸仙纪念医院

Dates

Publication Date: 20260512
Application Date: 20260128

Claims (12)

1. A method for predicting cardiovascular disease based on a large electrocardiographic model, the method being computer-implemented and comprising: Acquiring a pre-training electrocardio fragment set and a fine-tuning electrocardio fragment set, wherein the pre-training electrocardio fragment set comprises multi-lead electrocardio fragments of each subject in an unlabeled subject group, and the fine-tuning electrocardio fragment set comprises multi-lead electrocardio fragments of each subject in the labeled subject group; The method comprises the steps of inputting a pre-training electrocardio segment set into an electrocardio large model for self-supervision pre-training to obtain a pre-trained electrocardio large model, wherein the self-supervision pre-training comprises the steps of extracting features of any multi-lead electrocardio segment input to obtain a corresponding potential representation feature sequence, discretizing each potential representation feature vector in the potential representation feature sequence to obtain a corresponding discretization index sequence and a context target vector sequence, carrying out random masking on part of potential representation feature vectors in the potential representation feature sequence, and encoding the potential representation feature sequences after the random masking to obtain a corresponding prediction index probability distribution sequence and a context prediction vector sequence, so as to obtain a plurality of discretization index sequences, a plurality of prediction index probability distribution sequences, a plurality of context target vector sequences and a plurality of context prediction vector sequences corresponding to the pre-training electrocardio segment set, and minimizing a multitask loss function, wherein the multitask loss function comprises context comparison between the corresponding context target vector sequence and the context prediction vector sequence and context prediction vector sequence, and language modeling loss between the corresponding prediction index probability distribution sequence and the context prediction index sequence; Inputting the fine-tuning electrocardio segment set into a pre-trained electrocardio large model, and carrying out model fine tuning on the pre-trained electrocardio large model according to the output disease prediction result and the real disease label of the corresponding subject to obtain a cardiovascular disease prediction model; inputting the multi-lead electrocardio segments of the individual to be detected into the cardiovascular disease prediction model to obtain a corresponding cardiovascular disease prediction result.
2. The method of claim 1, wherein the multi-lead electrocardiographic fragment comprises two multi-lead electrocardiographic fragments; The self-supervision pre-training method comprises the steps of carrying out feature extraction on any input multi-lead electrocardio segment to obtain a corresponding potential representation feature subsequence, carrying out discretization processing on each potential representation feature sub-vector in the potential representation feature sub-sequence to obtain a corresponding discretization index sequence and a context target vector sequence, carrying out random masking on part of potential representation feature vectors in the potential representation feature sub-sequence, and encoding the potential representation feature sequences after the random masking to obtain a corresponding prediction index probability distribution sequence and a context prediction vector sequence, so as to obtain a plurality of discretization index sequences, a plurality of prediction index probability distribution sequences, a plurality of context target vector sequences and a plurality of context prediction vector sequences corresponding to the pre-training electrocardio segment set, and minimizing a multitask loss function, wherein the multitask loss function comprises context contrast loss between the corresponding context target vector sequence and the context prediction vector sequence, and masking language modeling loss between the corresponding prediction index probability distribution sequence and the discretization index sequence.
3. The method of claim 2, wherein the self-monitoring pre-training further comprises performing feature extraction for two multi-lead electrocardiographic segments of any one of the input multi-lead electrocardiographic segments to obtain corresponding two potential representation feature subsequences, constructing a positive sample pair using the two potential representation feature subsequences of the same multi-lead electrocardiographic segment in the pre-training electrocardiographic segment set, constructing a negative sample pair using the two potential representation feature subsequences between different multi-lead electrocardiographic segments in the pre-training electrocardiographic segment set, and minimizing the multi-task loss function further comprising a multi-segment contrast loss between the positive sample pair and the negative sample pair.
4. The method for predicting cardiovascular disease according to claim 1 or 3, wherein the electrocardiographic large model comprises a feature encoding module, a quantitative discrete module, a context characterization extraction module and a multi-task joint training module; The feature coding module is used for extracting potential representation feature vectors of the multi-lead electrocardio segments in a plurality of time steps to obtain the potential representation feature sequence; the quantization discrete module is used for carrying out discretization processing on each potential representation feature vector in the potential representation feature sequence to generate a corresponding discretization index sequence and a context target vector sequence; The context representation extraction module is used for carrying out random masking on partial potential representation feature vectors in the potential representation feature sequences, and encoding the potential representation feature sequences after the random masking to obtain corresponding prediction index probability distribution sequences and context prediction vector sequences; The multi-task combined training module is used for minimizing a multi-task loss function so as to obtain the pre-trained electrocardiographic large model.
5. The method according to claim 4, wherein the electrocardiographic large model further comprises a disease probability prediction module; After obtaining the cardiovascular disease prediction model based on the electrocardiographic large model, the disease probability prediction module is used for carrying out linear mapping on potential representation feature sequences corresponding to the multi-lead electrocardiographic fragments of the individual to be detected according to the learned weight matrix, and calculating a corresponding disease prediction result by utilizing a first nonlinear activation function.
6. The method of claim 4, wherein the feature encoding module comprises a plurality of sequentially stacked residual convolution blocks, the residual convolution blocks comprising a one-dimensional convolution layer, a normalization layer, and a second nonlinear activation function.
7. The method of claim 4, wherein the discretizing each potential representation feature vector in the sequence of potential representation features by using a plurality of codebooks comprises: Selecting most similar entries from the codebook by utilizing Gumbel-Softmax algorithm aiming at the potential representation feature vectors, wherein the codebook comprises a plurality of preset entries; splicing a plurality of most similar entries respectively obtained from the codebooks to obtain discretized indexes of the potential representation feature vectors; Obtaining a discretization index sequence of the potential representation feature sequence according to the discretization index of the potential representation feature vector; And carrying out linear transformation on each discretization index in the discretization index sequence to obtain a corresponding context target vector sequence.
8. The method of claim 4, wherein the context-characterization extraction module comprises a mask generator, an encoder, and a mapping module, the encoder comprising a plurality of sequentially stacked convolution-enhanced transducers; The mask generator is used for randomly selecting one potential representation feature vector in the potential representation feature sequence as a starting position, masking a plurality of subsequent continuous potential representation feature vectors according to a preset mask probability from the starting position, and obtaining a potential representation feature sequence after random masking; In the encoder, the potential representation feature sequence after the random mask is input into a first convolution enhancement converter for convolution and feature fusion to obtain an intermediate feature sequence, and the intermediate feature sequence is sent to a next convolution enhancement converter for convolution and feature fusion until the last convolution enhancement converter is used for convolution and feature fusion to obtain a corresponding context fusion sequence; the mapping module is used for mapping the context fusion sequence into a corresponding prediction index probability distribution sequence and a corresponding context prediction vector sequence.
9. The method of claim 8, wherein the convolution enhancement transformer comprises a depth convolution module and an attention module; the depth convolution module is used for convolving the potential representation feature sequence after the input random mask or the intermediate feature sequence received from the previous convolution enhancement converter to obtain a corresponding local feature enhancement sequence; The attention module is used for carrying out multistage attention fusion on the local feature enhancement sequence to obtain a corresponding intermediate feature sequence or context fusion sequence.
10. A cardiovascular disease prediction device based on an electrocardiographic large model, characterized in that the cardiovascular disease prediction device comprises: An acquisition module configured to acquire a pre-trained electrocardiograph fragment set comprising multi-lead electrocardiograph fragments for each subject in an unlabeled subject group and a fine-tuned electrocardiograph fragment set comprising multi-lead electrocardiograph fragments for each subject in a labeled subject group; The system comprises a pre-training module, a self-supervision pre-training module, a minimum multitask loss function, a random masking module, a context prediction vector sequence and a context prediction vector sequence, wherein the pre-training module is configured to input the pre-training electrocardio segment set into an electrocardio large model to perform self-supervision pre-training to obtain a pre-trained electrocardio large model, the self-supervision pre-training comprises performing feature extraction on any multi-lead electrocardio segment input to obtain a corresponding potential representation feature sequence, performing discretization processing on each potential representation feature vector in the potential representation feature sequence to obtain a corresponding discretization index sequence and a context target vector sequence, performing random masking on part of potential representation feature vectors in the potential representation feature sequence, and encoding the potential representation feature sequences after the random masking to obtain a corresponding prediction index probability distribution sequence and a context prediction vector sequence, so as to obtain a plurality of discretization index sequences, a plurality of prediction index probability distribution sequences, a plurality of context target vector sequences and a plurality of context prediction vector sequences corresponding to the pre-training electrocardio segment set; the model fine tuning module is configured to input the fine tuning electrocardio segment set into a pre-trained electrocardio large model, and perform model fine tuning on the pre-trained electrocardio large model according to the output disease prediction result and the real disease label of the corresponding subject to obtain a cardiovascular disease prediction model; The prediction module is configured to input the multi-lead electrocardio segments of the individual to be detected into the cardiovascular disease prediction model to obtain a corresponding cardiovascular disease prediction result.
11. An electronic device comprising a memory and a processor, wherein the memory is configured to store one or more computer instructions, wherein the one or more computer instructions are executed by the processor to implement the cardiovascular disease prediction method of any of claims 1-9.
12. A computer readable storage medium having stored thereon computer instructions which, when executed by a processor, implement the cardiovascular disease prediction method of any of claims 1 to 9.

Description

Cardiovascular disease prediction method and device based on electrocardiographic large model, electronic equipment and medium Technical Field The disclosure relates to the technical field of medical signal processing and artificial intelligence, in particular to a cardiovascular disease prediction method and device based on an electrocardiographic large model, electronic equipment and medium. Background Cardiovascular disease is one of the major health challenges worldwide. Electrocardiography is a widely acquired non-invasive physiological signal whose waveform contains key information reflecting the state of the heart. In recent years, electrocardiographic-based analysis and identification techniques have received widespread attention, but the existing methods still have significant limitations. At present, an electrocardiogram analysis method based on deep learning mostly adopts a full-supervision learning paradigm, and the performance of the method is seriously dependent on large-scale and high-quality manual annotation data. However, the special labeling cost of the electrocardiogram is high and the period is long, so that a large amount of acquired label-free electrocardiographic data cannot be effectively utilized. To reduce the dependence on annotation data, some studies have attempted to introduce self-supervised learning, learning characterizations from unlabeled data by designing pre-training tasks. However, existing self-supervision methods mostly operate directly on original signals or shallow features, such as local signal reconstruction or inter-sample comparison, and it is difficult to extract robust, high-order semantic features strongly related to heart states from the features. Such methods are susceptible to signal noise, individual differences and acquisition conditions, and have limited learned feature discrimination. Therefore, there is an urgent need for a method that can more fully utilize unlabeled electrocardiographic data, that is more discriminative and robust in learned features, and that can effectively support cardiovascular disease prediction. Disclosure of Invention In order to solve the problems in the related art, embodiments of the present disclosure provide a cardiovascular disease prediction method and apparatus, an electronic device, and a medium based on an electrocardiographic large model. In a first aspect, embodiments of the present disclosure provide a method for cardiovascular disease prediction based on a large electrocardiographic model, the method implemented by a computer, comprising: Acquiring a pre-training electrocardio fragment set and a fine-tuning electrocardio fragment set, wherein the pre-training electrocardio fragment set comprises multi-lead electrocardio fragments of each subject in an unlabeled subject group, and the fine-tuning electrocardio fragment set comprises multi-lead electrocardio fragments of each subject in the labeled subject group; The method comprises the steps of inputting a pre-training electrocardio segment set into an electrocardio large model for self-supervision pre-training to obtain a pre-trained electrocardio large model, wherein the self-supervision pre-training comprises the steps of extracting features of any multi-lead electrocardio segment input to obtain a corresponding potential representation feature sequence, discretizing each potential representation feature vector in the potential representation feature sequence to obtain a corresponding discretization index sequence and a context target vector sequence, carrying out random masking on part of potential representation feature vectors in the potential representation feature sequence, and encoding the potential representation feature sequences after the random masking to obtain a corresponding prediction index probability distribution sequence and a context prediction vector sequence, so as to obtain a plurality of discretization index sequences, a plurality of prediction index probability distribution sequences, a plurality of context target vector sequences and a plurality of context prediction vector sequences corresponding to the pre-training electrocardio segment set, and minimizing a multitask loss function, wherein the multitask loss function comprises context comparison between the corresponding context target vector sequence and the context prediction vector sequence and context prediction vector sequence, and language modeling loss between the corresponding prediction index probability distribution sequence and the context prediction index sequence; Inputting the fine-tuning electrocardio segment set into a pre-trained electrocardio large model, and carrying out model fine tuning on the pre-trained electrocardio large model according to the output disease prediction result and the real disease label of the corresponding subject to obtain a cardiovascular disease prediction model; inputting the multi-lead electrocardio segments of the individual to be det