CN-122004883-A - Method and device for extracting 12-lead electrocardiogram space-time combined features

CN122004883ACN 122004883 ACN122004883 ACN 122004883ACN-122004883-A

Abstract

The application provides an extraction method of 12-lead electrocardiogram space-time joint features, which comprises the steps of obtaining an original signal of a 12-lead electrocardiogram, preprocessing the original signal to obtain an input matrix, carrying out preliminary feature mapping on the input matrix to generate consistent input features, carrying out parallel feature extraction on the consistent input features to obtain time dimension features, lead space features and global convolution features, and carrying out dimension alignment and fusion on the time dimension features, the lead space features and the global convolution features to generate the space-time joint features. Effective information is comprehensively mined from three dimensions of time evolution, lead space and global morphology through three paths of parallel extraction, and key features are highlighted by combining dimensional alignment and refinement fusion strategies, so that the completeness and robustness of feature expression can be effectively improved, and the application requirements of clinical electrocardiographic inverse problem solving and accurate diagnosis and treatment are met.

Inventors

Zu lingyun
ZHAO PENGHUI
YIN ZHAOWEI
CAI JIAGENG
An hang
HU ANYI

Assignees

北京大学第三医院（北京大学第三临床医学院）

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. A method for extracting 12-lead electrocardiogram space-time joint features, the method comprising: Acquiring an original signal of a 12-lead electrocardiogram, and preprocessing the original signal to obtain an input matrix; performing preliminary feature mapping on the input matrix to generate consistent input features; Carrying out parallel feature extraction on the consistency input features to obtain time dimension features, lead space features and global convolution features; and carrying out dimension alignment and fusion on the time dimension feature, the lead space feature and the global convolution feature to generate a space-time joint feature.
2. The method of claim 1, wherein the step of obtaining and preprocessing the raw signals of the 12-lead electrocardiogram to obtain an input matrix comprises: Acquiring an original signal of a 12-lead electrocardiogram; sequentially performing power frequency and high frequency noise suppression and baseline drift correction processing on the original signal to obtain a correction signal; r wave detection is carried out on the correction signal to determine an R wave position, and the R wave position is used for cutting out a cardiac cycle signal; And carrying out voltage amplitude normalization processing on the cut cardiac cycle signals to obtain an input matrix with the size of 12×t, wherein 12 corresponds to 12 lead channels, and t is the sampling point length of the cardiac cycle.
3. The method of claim 1, wherein the step of performing a preliminary feature mapping on the input matrix to generate consistent input features comprises: determining dimension conversion parameters required by feature mapping according to the input matrix; performing feature conversion on the input matrix, and mapping the input matrix to a feature space with a preset dimension to obtain feature data; And carrying out unified standardization processing on the characteristic data to obtain consistent input characteristics.
4. The method of claim 1, wherein the step of parallel feature extraction of the consistent input features results in a time dimension feature, a lead spatial feature, and a global convolution feature, comprising: Starting three parallel extraction paths by taking the consistency input characteristics as shared input; extracting a time dimension feature from the consistency input feature by a time slice self-attention operation; extracting lead spatial features from the consistent input features through an inter-lead self-attention mechanism; extracting global convolution features from the consistent input features by multi-scale one-dimensional convolution; and synchronously outputting the time dimension characteristic, the lead space characteristic and the global convolution characteristic.
5. The method of claim 4, wherein the step of extracting a time dimension feature from the consistent input feature by a time slice self-attention operation comprises: dividing the consistency input features into overlapping time segments according to preset length and step length, generating query vectors, key vectors and value vectors through linear mapping, calculating similarity among vectors, normalizing to obtain attention weights, weighting and summing the value vectors based on the weights, and outputting time dimension features.
6. The method of claim 4, wherein the step of extracting lead spatial features from the consistent input features via an inter-lead self-attention mechanism comprises: rearranging the consistency input features into a lead sequence with the length of 12, encoding the lead sequence into lead-level feature vectors, generating query vectors, key vectors and value vectors through linear mapping, calculating the similarity of inner products among the vectors, normalizing the similarity to obtain the attention weights among the leads, weighting and summing the value vectors based on the weights, and outputting lead space features.
7. The method of claim 1, wherein the step of dimensionally aligning and fusing the time dimension feature, the lead spatial feature, and the global convolution feature to generate a spatio-temporal joint feature further comprises: Performing dimension matching processing on the time dimension feature, the lead space feature and the global convolution feature to enable the time dimension feature, the lead space feature and the global convolution feature to be in the same feature dimension; fusing the time dimension characteristics, the lead space characteristics and the global convolution characteristics after dimension alignment by adopting a residual error addition or weighted combination mode to obtain a fusion result; And optimizing the fusion result through nonlinear transformation to generate space-time joint characteristics with time evolution rules, lead space coupling relations and global morphological information.
8. An extraction device for 12-lead electrocardiogram time-space joint features, which is characterized in that the extraction device for 12-lead electrocardiogram time-space joint features is used for realizing the steps of the extraction method for 12-lead electrocardiogram time-space joint features according to any one of claims 1 to 7, Comprising the following steps: The acquisition module is used for acquiring an original signal of the 12-lead electrocardiogram and preprocessing the original signal to obtain an input matrix; The preliminary feature mapping module is used for carrying out preliminary feature mapping on the input matrix to generate consistent input features; The feature extraction module is used for carrying out parallel feature extraction on the consistency input features to obtain time dimension features, lead space features and global convolution features; and the fusion module is used for carrying out dimension alignment on the time dimension feature, the lead space feature and the global convolution feature and fusing to generate a space-time joint feature.
9. An apparatus for extracting 12-lead electrocardiogram spatiotemporal joint features, comprising a processor, a memory and a computer program stored on the memory and executable on the processor, the computer program when executed by the processor implementing the method for extracting 12-lead electrocardiogram spatiotemporal joint features according to any of claims 1 to 7.
10. A computer readable storage medium, wherein a computer program is stored on the computer readable storage medium, which when executed by a processor implements the method for extracting spatiotemporal joint features of 12-lead electrocardiography according to any one of claims 1 to 7.

Description

Method and device for extracting 12-lead electrocardiogram space-time combined features Technical Field The invention relates to the technical field of biomedical signal processing, in particular to a method and a device for extracting 12-lead electrocardiogram space-time combined features. Background The myocardial infarction is used as a cardiovascular emergency seriously threatening the life health of human beings, and the accurate positioning of the infarcted area is a core premise of clinical risk stratification, blood circulation reconstruction strategy formulation and arrhythmia ablation scheme design, so that the treatment effect and the prognosis of patients are directly influenced. In the existing clinical detection technology, although the late gadolinium enhanced cardiac magnetic resonance (LGE-CMR) can intuitively present the form and range of infarct scar, is a main method for evaluating myocardial infarction at present, is limited by the defects of high examination cost, strict requirement on the breath-holding capacity of patients, need of injection of contrast agent, potential influence on renal function and the like, and is difficult to be used as a conventional screening or bedside rapid evaluation tool. In contrast, the standard 12-lead Electrocardiogram (ECG) is a preferred means for the primary diagnosis and evaluation of clinical myocardial infarction by virtue of the outstanding advantages of convenient operation, low cost, no wound, no radiation and the like. However, the 12-lead ECG is used as the body surface indirect reflection of the heart electrical activity, the spatial resolution is naturally limited, only the electric signal change of the limited sites of the body surface can be captured, the electrophysiological distribution of the ventricular surface is difficult to precisely map, the infarct range and the lesion degree of different coronary blood supply areas cannot be finely distinguished, and the clinical precise diagnosis and treatment are challenged. In order to break through the bottleneck of insufficient spatial resolution of 12-lead ECG, a series of data-driven electrocardiographic inverse problem (ECGI) solving methods are proposed in the academic circles and the industry, and a series of data-driven electrocardiographic inverse problem (ECGI) solving methods attempt to reconstruct transmembrane potential (TMP) from body surface lead signals to ventricular surfaces by mining deep information in electrocardiographic signals so as to infer an infarct area. However, the existing method still has significant defects in practical application, and the improvement of positioning precision and stability is restricted: First, effective information mining is inadequate. The preprocessing of electrocardiosignals in the prior art is mostly limited to conventional operations such as basic filtering, baseline correction, amplitude normalization and the like, and feature extraction focuses on a small amount of traditional dominant indexes such as ST elevation, voltage peak value, QRS wave group time limit and the like, and time sequence evolution rules (such as dynamic processes of myocardial cell depolarization and repolarization) hidden in complete electrocardiosignal waveforms and space correlation characteristics among multiple leads cannot be deeply mined, so that feature information dimension of an input inverse problem solving model is single, expression capability is insufficient, and stability and positioning accuracy of inverse problem solving are directly affected. And secondly, redundant information is not effectively processed. There are a large number of highly correlated waveform components between the different leads of the 12-lead ECG, reflecting the correlation of the cardiac electrical activity on the body surface projections. However, in the prior art, in the feature construction process, multi-lead data is often processed by adopting a simple splicing or equal-weight superposition mode, and a targeted compression and screening mechanism for redundant information or low-value information is lacked, so that the extracted features contain a large number of invalid redundant components, the calculation complexity of a model is increased, noise interference is possibly amplified, and the positioning stability of myocardial infarction areas is weakened. Thirdly, the multidimensional information fusion strategy is rough. The time dynamic characteristics, the lead spatial characteristics and the global rhythm morphological characteristics contained in the electrocardiosignal are mutually related and cooperatively act to jointly support the accurate judgment of the infarcted area. However, in the existing method, rough fusion modes such as simple splicing, direct summation and the like are mostly adopted, and a refined fusion design is lacked, so that redundant features are not fully restrained, key features related to diagnosis cannot be effectively highlighted, and