Search

CN-121999077-A - Physiological constraint type ECG reconstruction method and system based on multi-branch gating network

CN121999077ACN 121999077 ACN121999077 ACN 121999077ACN-121999077-A

Abstract

The invention discloses a physiological constraint type ECG reconstruction method and system based on a multi-branch gating network, and relates to the technical field of biomedical signal processing and deep learning. The method adopts a multi-encoder-multi-decoder attention-gated U-shaped network (MED-AG-UNet), takes I, II and V3 three-lead electrocardiosignals as input, independently extracts characteristics through the multi-encoder to avoid interference, realizes intelligent information fusion by using a cross attention mechanism, screens key characteristics by means of attention-gated jump connection, finally generates eight basic leads in parallel by the multi-decoder, and calculates the remaining four limb leads by combining a physiological constraint module based on Einthoven law and Goldberger equation, thereby reconstructing a complete standard 12-lead electrocardiogram with high fidelity. The invention avoids the characteristic interference among leads, ensures that the reconstructed signals accord with physiological rules, and realizes the reconstruction of few leads with high fidelity and high precision.

Inventors

  • LIU HUAZHU
  • LIANG PEILIN
  • TANG SUIGU
  • ZHAO XIAOFANG
  • LIN SHENGXIN

Assignees

  • 东莞理工学院

Dates

Publication Date
20260508
Application Date
20260113

Claims (8)

  1. 1. A physiological constraint type ECG reconstruction method based on a multi-branch gating network, comprising the steps of: s1, inputting the preprocessed lead I, II and V3 electrocardiosignals into three independent one-dimensional convolutional encoders respectively, and extracting deep bottleneck feature vectors and multi-scale jump feature graphs of the leads in parallel; S2, conducting cross-linked attention fusion, namely constructing a global context space by utilizing a cross-attention module, conducting cross-linked intelligent fusion on the exclusive morphological characteristics through a learnable query vector, and generating decoding guide characteristics aiming at different target leads: (1) Wherein, the Q is a learnable query vector corresponding to each target lead, K is different input lead characteristics, V is a key and value matrix obtained by linear transformation of the global context information, and d and K are feature dimensions; s3, parallel decoding and feature gating, namely reconstructing lead I, II and V1 to V6 signals respectively through eight parallel decoders based on decoding guide features generated in the S2, introducing an attention gating module in the up-sampling process of each stage of the decoder, taking the features of each stage of decoder as gating signals, and carrying out self-adaptive weighted screening on jump feature graphs of corresponding layers of the encoder, wherein the calculation process of the attention gating is defined by the following formula: ; Wherein, the In order to have an intermediate fusion feature, And For the convolution weights to be the same, In order for the offset to be a function of, In order to output the transform convolution, As a function of the Sigmoid, In order to generate the attention coefficient(s), Gating features for final output; S4, physiological constraint synthesis, namely inputting the lead I and II signals preprocessed by the S1 into a physiological constraint module, and calculating the remaining four limb leads according to the linear relation of the electrocardio vectors, wherein the calculation formula is as follows: lead III = reconstruction value of lead II-reconstruction value of lead I; Lead avr= -0.5× (reconstruction value of lead I + reconstruction value of lead II); Lead aVL = reconstruction value for lead I-0.5 x reconstruction value for lead II; lead aVF = reconstruction value for lead II-0.5 x reconstruction value for lead I; s5, outputting a complete lead, namely combining eight lead signals reconstructed in the step S3 with four lead signals calculated in the step S4, and outputting a complete standard 12-lead electrocardiogram.
  2. 2. The physiological constraint type ECG reconstruction method based on the multi-branch gating network as set forth in claim 1, wherein the preprocessing in S1 specifically comprises the steps of filtering an original electrocardiosignal by adopting a fourth-order Butterworth band-pass filter with a passband of 0.5Hz to 40Hz, resampling all signals to a sampling rate of 250Hz, uniformly cutting off or filling the signals to 2048 sampling points, and finally carrying out maximum-minimum normalization processing.
  3. 3. The physiological constraint type ECG reconstruction method based on the multi-branch gating network as set forth in claim 1, wherein the one-dimensional convolution encoder and decoder are each composed of a plurality of double convolution units in a stacked mode, the double convolution units comprise two serially connected one-dimensional convolution blocks, and each convolution block sequentially comprises a one-dimensional convolution layer, a batch normalization layer, a Dropout layer and a LeakyReLU activation function layer.
  4. 4. The method of claim 1, wherein in S2, eight independent learnable query vectors Q are initialized for the eight target-lead decoders, respectively.
  5. 5. The method of claim 1, wherein S3, each decoder receives multi-scale jump features from all three encoders via jump connections and performs screening via independent attention gating modules.
  6. 6. The physiological constraint type ECG reconstruction method based on the multi-branch gating network as set forth in claim 1, wherein: the method uses PTB-XL clinical electrocardiographic datasets for model training, validation and testing.
  7. 7. A physiological constraint type ECG reconstruction system based on a multi-branch gating network, which is used for realizing the method of any one of claims 1-6, and comprises a preprocessing module, a multi-encoder module, a cross attention fusion module, a multi-decoder module, a physiological constraint calculation module and an output module, wherein the preprocessing module is used for carrying out standardized processing on input I, II and V3 lead signals, the multi-encoder module is used for extracting exclusive characteristics of each input lead, the cross attention fusion module is used for carrying out attention calculation and information fusion of S2, the multi-decoder module is used for reconstructing eight basic leads in parallel and integrating the attention gating sub-module to execute characteristic screening of S3, the physiological constraint calculation module is used for executing linear equation calculation of S4, and the output module is used for synthesizing and outputting a complete 12-lead electrocardiogram.
  8. 8. The physiological constraint ECG reconstruction system based on a multi-branch gating network as set forth in claim 7, wherein said multi-encoder module, cross-attention fusion module, multi-decoder module and internal attention gating submodule together form an end-to-end deep learning model which is jointly optimized for training by loss function supervision.

Description

Physiological constraint type ECG reconstruction method and system based on multi-branch gating network Technical Field The invention relates to the technical field of biomedical signal processing and deep learning, in particular to a physiological constraint type ECG reconstruction method and system based on a multi-branch gating network. Background More than 3 billion electrocardiographic examinations are performed worldwide each year, and 12-lead electrocardiography has become the basic diagnosis for assessing cardiovascular disease. Using 10 separate arrays of skin surface electrodes, a series of 12 different electrical signals were acquired to aid in the diagnosis of various cardiopulmonary diseases. Despite advances in electrocardiographic vector diagrams and other recording techniques, including the Mason-Likar system, most clinical diagnoses still rely on standard 12-lead electrocardiographs, the acquisition of which is highly dependent on the expertise of the operator. Special equipment that can only be used by hospitals or clinics is required, as well as specially trained personnel to perform and interpret electrocardiograms. Over the past few years, advances in technology have made it possible to monitor specific cardiac activity with higher quality and speed through wearable devices, including smart watches, patch monitors, and applications. However, in this case, electrocardiographic monitoring is often limited to a single lead (typically an I-lead) or several limb leads, which is insufficient to continuously diagnose abnormalities limited to a specific myocardial area only, such as acute myocardial infarction. Because the particular pattern of acute MI may be reflected in limb leads, chest leads, or a combination of limb leads and chest leads, current guidelines require clinical interpretation using a 12-lead standard electrocardiogram. The 12 leads in standard electrocardiograms are not completely independent and are known to be partially correlated, so in the last 30 years techniques have been proposed to synthesize a completely standardized limited set of ECG leads. While the initial evolution in this area has relied on linear transformation models, the popularity of Artificial Intelligence (AI) has led to the development of more complex approaches. Previous studies have relied primarily on patient-specific models, or are derived from limited data sets, which may limit their popularity. Therefore, there is a need to design a new physiological constraint type ECG reconstruction method and system based on a multi-branch gating network to solve the above problems. Disclosure of Invention The invention provides a physiological constraint type ECG reconstruction method and system based on a multi-branch gating network, which aim to solve at least one technical problem in the prior art. In order to achieve the above purpose, the present invention adopts the following technical scheme: A physiological constraint type ECG reconstruction method based on a multi-branch gating network, comprising the steps of: s1, inputting the preprocessed lead I, II and V3 electrocardiosignals into three independent one-dimensional convolutional encoders respectively, and extracting deep bottleneck feature vectors and multi-scale jump feature graphs of the leads in parallel; S2, conducting cross-linked attention fusion, namely constructing a global context space by utilizing a cross-attention module, conducting cross-linked intelligent fusion on the exclusive morphological characteristics through a learnable query vector, and generating decoding guide characteristics aiming at different target leads: (1) Wherein, the Q is a learnable query vector corresponding to each target lead, K is different input lead characteristics, V is a key and value matrix obtained by linear transformation of the global context information, and d and K are feature dimensions; s3, parallel decoding and feature gating, namely reconstructing lead I, II and V1 to V6 signals respectively through eight parallel decoders based on decoding guide features generated in the S2, introducing an attention gating module in the up-sampling process of each stage of the decoder, taking the features of each stage of decoder as gating signals, and carrying out self-adaptive weighted screening on jump feature graphs of corresponding layers of the encoder, wherein the calculation process of the attention gating is defined by the following formula: ; Wherein, the In order to have an intermediate fusion feature,AndFor the convolution weights to be the same,In order for the offset to be a function of,In order to output the transform convolution,As a function of the Sigmoid,In order to generate the attention coefficient(s),Gating features for final output; S4, physiological constraint synthesis, namely inputting the lead I and II signals preprocessed by the S1 into a physiological constraint module, and calculating the remaining four limb leads according to the