Method for detecting QRS complex of electrocardiogram signal, apparatus therefor, device and medium

Patent metadata

Field	Value
Publication number	US-12171564-B1
Application number	US-202418818084-A
Country	US
Kind code	B1
Filing date	Aug 28, 2024
Priority date	Jan 24, 2024
Publication date	Dec 24, 2024
Grant date	Dec 24, 2024

Abstract

A method for detecting a QRS complex of an electrocardiogram signal, an apparatus therefor, a device and a medium are provided. A QRS complex of an electrocardiogram signal is detected by a detection model including a DenseNet and a dual-channel Long Short-Term Memory (LSTM). Spatial feature information of the QRS complex in the electrocardiogram signal is learned through the DenseNet, and then the spatial feature learned by the DenseNet and the time sequence data are input into respective channels of the dual-channel LSTM, so that the dual-channel LSTM can fuse the spatial information and the time sequence information of the QRS complex in the electrocardiogram signal, thereby improving a segmentation effect of the model on features, and finally improving accuracy of detecting the QRS complex.

First claim

What is claimed is:

1. A method for detecting a QRS complex of an electrocardiogram signal, comprising:
acquiring time sequence data of the electrocardiogram signal from an electrocardiogram signal;
constructing a detection model consisting of a DenseNet and a dual-channel Long Short-Term Memory (LSTM), inputting the time sequence data into the detection model to extract a spatial feature of the electrocardiogram signal from the time sequence data according to the DenseNet, inputting the time sequence data and the spatial feature into respective channels of the dual-channel LSTM, and detecting a QRS complex of the electrocardiogram signal by the dual-channel LSTM;
wherein the inputting the time sequence data and the spatial feature into respective channels of the dual-channel LSTM, and detecting a QRS complex of the electrocardiogram signal by the dual-channel LSTM, comprises:
inputting the time sequence data into a first channel of the dual-channel LSTM to obtain a first feature output by the first channel, and inputting the spatial feature into a second channel of the dual-channel LSTM to obtain a second feature output by the second channel;
fusing the first feature and the second feature by means of respective attention weights by using an attention mechanism to obtain a third feature; wherein a sum of attention weights of two channels is equal to 1;
detecting the QRS complex according to the third feature;
wherein the detection model is trained by preprocessed training data; and preprocessing of training data comprises:
normalizing the training data;
carrying out sample equalization for various labels on normalized training data;
carrying out data enhancement on the training data after sample equalization;

wherein in training process, the detection model obtains the attention weights by following equation:
Attention( H LSTM1 ,H LSTM2 )=softmax( W ·tanh( U·H LSTM1 +V·H LSTM2 ));

wherein H LSTM1 is an output feature of the first channel, H LSTM2 is an output feature of the second channel, tanh is a hyperbolic tangent function, softmax( ) is an activation function, W, U, V are learnable weight parameters, and Attention(H LSTM1 ,H LSTM2 ) is the attention weights;
wherein in the training process, the detection model fuses the output feature of the first channel and the output feature of the second channel by following equation:
Fused Feature=Attention( H LSTM1 ,H LSTM2 )· H LSTM1 +(1−Attention( H LSTM1 ,H LSTM2 ))· H LSTM2 ;

wherein Fused Feature is a feature obtained by fusing the output feature of the first channel and the output feature of the second channel.

2. The method according to claim 1 , wherein the carrying out sample equalization for various labels on normalized training data comprises:
repeating following steps until a predetermined data equalization degree is reached; the following steps comprise:
selecting minority category samples in the normalized training data;
randomly selecting a corresponding neighbor for each of the minority category samples, wherein the neighbor is one of several nearest neighbor samples of each of the minority category samples;
generating a new sample between each of the minority category samples and corresponding neighbor by a linear interpolation method.

3. The method according to claim 1 , wherein a loss function of the detection model in the training process is:
L total =L cls +λL reg ;
L cls ( y cls ,ŷ cls )=−( y cls ·log( ŷ cls )+(1− y cls )·log(1− ŷ cls ));

L reg ( t reg ,{circumflex over (t)} reg )=Smooth L 1( t reg −{circumflex over (t)} reg );
where L total denotes a loss function of the detection model, L cls denotes a classification loss, L reg denotes a regression loss, λ denotes a predetermined weight parameter, ŷ cls denotes the QRS complex detected by the detection model, y cls denotes a real category label, {circumflex over (t)} reg denotes a position of the QRS complex detected by the detection model, and t reg denotes a real position label.

4. The method according to claim 1 , wherein the training data is obtained by following steps:
using a wearable electrocardiogram data acquisition device to acquire an original electrocardiogram signal of a user;
using a high-pass filter to filter the original electrocardiogram signal;
downsampling the original electrocardiogram signal filtered to obtain the training data.

5. An electronic device, comprising at least one control processor and a memory communicated with the at least one control processor; wherein instructions executable by the at least one control processor are stored in the memory, and the instructions are executed by the at least one control processor, so that the at least one control processor is capable of executing the method for detecting the QRS complex of the electrocardiogram signal according to claim 1 .

6. The electronic device according to claim 5 , the instructions are executed by the at least one control processor, so that the at least one control processor is capable of executing following steps:
repeating following steps until a predetermined data equalization degree is reached; the following steps comprise:
selecting minority category samples in the normalized training data;
randomly selecting a corresponding neighbor for each of the minority category samples, wherein the neighbor is one of several nearest neighbor samples of each of the minority category samples;
generating a new sample between each of the minority category samples and corresponding neighbor by a linear interpolation method.

7. The electronic device according to claim 5 , wherein a loss function of the detection model in the training process is:
L total =L cls +λL reg ;
L cls ( y cls ,ŷ cls )=−( y cls ·log( ŷ cls )+(1− y cls )·log(1− ŷ cls ));

L reg ( t reg ,{circumflex over (t)} reg )=Smooth L 1( t reg −{circumflex over (t)} reg );
where L total denotes a loss function of the detection model, L cls denotes a classification loss, L reg denotes a regression loss, λ denotes a predetermined weight parameter, ŷ cls is denotes the QRS complex detected by the detection model, y cls denotes a real category label, {circumflex over (t)} reg denotes a position of the QRS complex detected by the detection model, and t reg denotes a real position label.

8. The electronic device according to claim 5 , wherein the training data is obtained by following steps:
using a wearable electrocardiogram data acquisition device to acquire an original electrocardiogram signal of a user;
using a high-pass filter to filter the original electrocardiogram signal;
downsampling the original electrocardiogram signal filtered to obtain the training data.

9. A non-transitory computer-readable storage medium, wherein computer-executable instructions are stored in the non-transitory computer-readable storage medium, and the computer-executable instructions are used to cause a computer to execute the method for detecting the QRS complex of the electrocardiogram signal according to claim 1 .

10. The non-transitory computer-readable storage medium according to claim 9 , the computer-executable instructions are used to cause a computer to execute following steps:
repeating following steps until a predetermined data equalization degree is reached; the following steps comprise:
selecting minority category samples in the normalized training data;
randomly selecting a corresponding neighbor for each of the minority category samples, wherein the neighbor is one of several nearest neighbor samples of each of the minority category samples;
generating a ne

Assignees

• Univ Central South

Inventors

• Guo Lin
• Ma Nan
• An Ying
• Zhan Qianyun
• Long Jun

Classifications

• A61B5/7225

  Details of analogue processing, e.g. isolation amplifier, gain or sensitivity adjustment, filtering, baseline or drift compensation (input circuits for detecting, measuring, or recording bioelectric or biomagnetic signals A61B5/30; specific diagnostic methods using bioelectric or biomagnetic signals A61B5/316) · CPC title

• A61B5/256

  Wearable electrodes, e.g. having straps or bands · CPC title

• A61B5/349 Primary

  Detecting specific parameters of the electrocardiograph cycle · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title