US 11,375,942 B2 · Method for extracting f wave from single-lead electrocardiography signal

Abstract

A method for extracting f wave from a single-lead electrocardiography signal is provided. The method includes: establishing a two-channel temporal convolutional neural network model, including two coding submodules, two mask estimation submodules, an information fusion module and two decoding submodules; constructing a training data set of ECG signals; training the two-channel temporal convolutional neural network model, and saving the trained model. The two-channel temporal convolution neural network encodes and decodes a QRST complex and the f wave respectively, and uses a mask of information fusion to estimate and construct regular items to restrict a distribution difference of QRST component coding features, so as to reduce the distortion of the QRST complex, select potential features of a QRST complex and f wave with high signal-to-noise ratio, and thus improve a detection accuracy of the f wave.

A single-lead f wave extraction method, comprising:

establishing a two-channel temporal convolutional neural network model, wherein the two-channel temporal convolutional neural network model comprises two coding submodules, two mask estimation submodules, an information fusion module and two decoding submodules;
constructing a training data set of electrocardiography (ECG) signals, training the two-channel temporal convolutional neural network model according to the training data set of ECG signals to obtain a trained model, and saving the trained model;
and
extracting a reconstructed time domain signal of an f wave by inputting a measured mixed ECG signal into the trained model;
wherein the two coding submodules are respectively configured to extract a coding feature vector of a ventricular QRST complex and a coding feature vector of the f wave from the measured mixed ECG signal;
the two mask estimation submodules are respectively configured to extract a potential feature vector of the f wave and a potential feature vector of the ventricular QRST complex according to the coding feature vector of the ventricular QRST complex and the coding feature vector of the f wave;
the information fusion module is configured to perform feature fusion on the potential feature vector of the f wave and the potential feature vector of the ventricular QRST complex in a coding space, to estimate a mask feature vector of the ventricular QRST complex and a mask feature vector of the f wave;
and the two decoding submodules are respectively configured to reconstruct a time domain signal of the ventricular QRST complex and a time domain signal of the f wave according to a weighted result of the mask feature vector and the coding feature vector of the ventricular QRST complex and a weighted result of the mask feature vector and the coding feature vector of the f wave to obtain a reconstructed time domain signal of the ventricular QRST complex and the reconstructed time domain signal of the f wave;
wherein the two-channel temporal convolutional neural network model is established based on a probability graph model, and the probability graph model is formulated by a probability factorization formula expressed as follows:
P
⁡
(
x
V
⁢
A
,
x
A
⁢
A
|
x
)
=
∫
x
⁢
P
⁡
(
x
V
⁢
A
|
Z
~
V
⁢
A
)
⁢
P
⁡
(
x
A
⁢
A
|
Z
~
A
⁢
A
)
⁢
P
⁡
(
Z
~
V
⁢
A
|
M
V
⁢
A
,
Z
V
⁢
A
)
·
⁢
⁢
P
⁡
(
Z
~
A
⁢
A
|
M
A
⁢
A
,
Z
A
⁢
A
)
⁢
P
⁡
(
M
V
⁢
⁢
A
|
F
V
⁢
A
,
F
A
⁢
A
)
⁢
P
⁡
(
M
A
⁢
A
|
F
V
⁢
A
,
F
A
⁢
A
)
·
⁢
⁢
P
⁡
(
F
V
⁢
A
|
Z
V
⁢
A
)
⁢
P
⁡
(
F
A
⁢
A
|
Z
A
⁢
A
)
⁢
P
⁡
(
Z
V
⁢
A
|
x
)
⁢
P
⁡
(
Z
A
⁢
A
|
x
)
⁢
dx
wherein P(xVA, xAA|x) represents a conditional distribution probability being decomposed in the probability factorization formula, x represents the measured mixed ECG signal, XVA represents the reconstructed time domain signal of the ventricular QRST complex, XAA represents the reconstructed time domain signal of the f wave, ZVA and ZAA respectively represent the coding feature vectors of the ventricular QRST complex and the f wave, FAA and FVA respectively represent the potential feature vectors of the ventricular QRST complex and the f wave, MVA and MAA represent the mask feature vectors of the ventricular QRST complex and the f wave, {tilde over (Z)}VA and {tilde over (Z)}AA represent weighted feature vectors, and P ( ) represents a probability.

190 more elements

Method for extracting f wave from single-lead electrocardiography signal

Life of this patent

Method for extracting f wave from single-lead electrocardiography signal

Life of this patent

Abstract

Claims9