Resonance sparse decomposition-based twelve-lead electrocardiosignal atrial fibrillation detection method
阅读说明：本技术 一种基于共振稀疏分解的十二导联心电信号房颤检测方法 (Resonance sparse decomposition-based twelve-lead electrocardiosignal atrial fibrillation detection method ) 是由 舒明雷 马昊 朱清 王英龙 陈超 刘辉 高天雷 周书旺 谢小云 孔祥龙 于 2020-03-27 设计创作，主要内容包括：一种基于共振稀疏分解的十二导联心电信号房颤检测方法,根据房颤信号的特点,利用共振稀疏分解对心电信号进行处理,将分解出的低共振分量部分送入神经网络中进行训练,逐个导联训练后通过投票算法(Voting)将各训练模型得出的概率进行联合运算,得到最终的检测结果,用于房颤检测时无需额外手工提取其他特征,网络结构简单,缩短了运算时间,可以实现实时房颤信号检测。针对房颤信号中“P波消失,出现F波”的表现特点,以共振稀疏分解为基础,突出房颤信号特点,通过简单的神经网络结构,保准准确率的基础上减少了运算时间。(A twelve-lead electrocardiosignal atrial fibrillation detection method based on resonance sparse decomposition is characterized in that according to the characteristics of atrial fibrillation signals, resonance sparse decomposition is utilized to process the electrocardiosignals, decomposed low-resonance component parts are sent into a neural network for training, probability obtained by training models is subjected to combined operation through a Voting algorithm (Voting) after one-by-one lead training, a final detection result is obtained, additional manual extraction of other characteristics is not needed when the method is used for detecting atrial fibrillation, the network structure is simple, operation time is shortened, and real-time atrial fibrillation signal detection can be realized. Aiming at the performance characteristics of 'P wave disappears and F wave appears' in the atrial fibrillation signal, the characteristics of the atrial fibrillation signal are highlighted on the basis of resonance sparse decomposition, and the operation time is reduced on the basis of accuracy guarantee through a simple neural network structure.)
1. A twelve-lead electrocardiosignal atrial fibrillation detection method based on resonance sparse decomposition is characterized by comprising the following steps:
a) slicing and drying the static twelve-lead electrocardiosignal y by using a computer to obtain an electrocardiosignal y' with fixed length and part of noise removed;
b) processing each lead of the electrocardiosignal y' with partial noise removed by a resonance sparse decomposition technology one by one to obtain a high-resonance signal component yHLow resonance signal component yLAnd a residual signal component n;
c) the decomposed low-resonance signal component y of each lead of the electrocardiosignal yLRespectively sending the signals into a neural network for training, and outputting a detection result after the training is finished;
d) and combining the detection results output by the leads, calculating the final classification result by adopting a Voting algorithm, and giving different weights to the leads to obtain the final detection result.
2. The resonance sparse decomposition-based twelve-lead cardiac signal atrial fibrillation detection method according to claim 1, wherein the step a) is implemented by the following steps:
a-1) slicing a static twelve-lead electrocardiosignal y by using a computer, and intercepting N sampling points in front of each lead to obtain a sliced electrocardiosignal;
a-2) processing the sliced electrocardio data by utilizing a 0.5-40Hz band-pass filter to obtain an electrocardio signal y' with partial noise removed.
3. The resonance sparse decomposition-based twelve-lead cardiac signal atrial fibrillation detection method according to claim 1, wherein the step b) is implemented by the following steps:
b-1) according to the formula y ═ yH+yL+ n decomposing the partial noise eliminated electrocardiosignal y' based on the resonance sparse decomposition algorithm, wherein yHIs a high resonance signal component, yLIs a low resonance signal component, n is a residual signal component; b-2) by the formula y ═ W1M1+W2M2+ n calculating partial noise eliminated electrocardiosignal y' with W1Is a high resonance signal component yHIn redundant dictionary M1Transform coefficient of down, W2Is a low resonance signal component yLIn redundant dictionary M2A lower transform coefficient, n being a residual signal component;
b-3) by the formula J (W)1,W2)＝||y′-W1M1-W2M2||2+λ1||W1||1+λ2||W2||2Calculating an objective function J (W) for measuring sparsity1,W2) In the formula of1And λ2As a weight coefficient, M1Is a high resonance signal component yHRedundant dictionary of, M2Is a low resonance signal component yLRedundant dictionaries of (1);
b-4) applying an iterative algorithm to the target function J (W)1,W2) Iteration is carried out to obtain an objective function J (W)1,W2) Using the objective function J (W)1,W2) The minimum value of (A) is obtained as a transformation coefficient W of the high-resonance signal component at that time1 *And a transformation coefficient W of a low resonance signal component2 *By the formula yH＝W1 *M1Will result in a high resonance signal component yHIn redundant dictionary M1Is represented by a new coefficient of variation, represented by the formula yL＝W2 *M2Will result in a low resonance signal component yLIn redundant dictionary M2Represented by the new coefficient of variation.
4. The resonance sparse decomposition-based twelve-lead cardiac signal atrial fibrillation detection method according to claim 1, wherein the step c) is implemented by the following steps:
the decomposed low resonance component y of the electrocardiosignalLPerforming network training as input of a neural network model, wherein the neural network model consists of four volume blocks, a bidirectional long-time and short-time memory network and a full connection layer, the bidirectional long-time and short-time memory network is connected behind the four volume blocks, the output dimensionality of the network is set to be 32, the network is converted into a characteristic value x with the output dimensionality of 2 through the full connection layer, and the probability P that a signal is normal is predicted through the characteristic value xNiAnd probability of atrial fibrillation PAFiWherein i is the serial number of different leads of the electrocardiogram data.
5. The resonance sparsity decomposition-based twelve-lead cardiac signal atrial fibrillation detection method according to claim 4, wherein the step d) is implemented by the following steps:
the probability of outputting twelve leads is according to a formulaAndcalculating to obtain final detection result, wherein P isNi+PAFi＝1，ωNiThe normal probability weight, ω, corresponding to the ith leadAFiThe probability weight of atrial fibrillation corresponding to the ith lead.
6. The resonance sparse decomposition-based twelve-lead cardiac signal atrial fibrillation detection method according to claim 2, characterized in that: the value of N in the step a-1) is 4096.
7. The resonance sparse decomposition-based twelve-lead cardiac signal atrial fibrillation detection method according to claim 4, wherein the detection method comprises the following steps: each convolution block of the four convolution blocks comprises a convolution layer and a pooling layer, the convolution layer is a one-dimensional convolution layer, the size of a convolution kernel is set to be 1 x 31, the step size is set to be 1, the padding is set to be 15, the dimensionalities of the convolution layers in the four convolution blocks are 8/64/64/8 respectively, the size of the kernel in the pooling layer is set to be 2, and the step size is set to be 2.
Technical Field
The invention relates to the technical field of electrocardiosignal classification detection, in particular to a twelve-lead electrocardiosignal atrial fibrillation detection method based on resonance sparse decomposition.
Background
The rapid detection of atrial fibrillation signals is a key problem in the processing of electrocardiosignals. At present, the commonly used atrial fibrillation detection algorithm mainly detects atrial fibrillation by analyzing P-wave dispersion or whether F-waves exist or not and analyzing RR interphase sequences, mostly adopts single-lead data in the detection process, is low in accuracy and is high in operation difficulty. With the development of deep learning, although the atrial fibrillation detection algorithm based on deep learning obtains a higher result, the deep neural network is mostly adopted, so that the calculation complexity is high, the calculation time is long, and the algorithm cannot be used for real-time atrial fibrillation detection.
Disclosure of Invention
In order to overcome the defects of the technology, the invention provides the twelve-lead electrocardiosignal atrial fibrillation detection method based on resonance sparse decomposition, which has high detection result accuracy.
The technical scheme adopted by the invention for overcoming the technical problems is as follows:
a twelve-lead electrocardiosignal atrial fibrillation detection method based on resonance sparse decomposition comprises the following steps:
a) slicing and drying the static twelve-lead electrocardiosignal y by using a computer to obtain an electrocardiosignal y' with fixed length and part of noise removed;
b) processing each lead of the electrocardiosignal y' with partial noise removed by a resonance sparse decomposition technology one by one to obtain a high-resonance signal component yHLow resonance signal component yLAnd a residual signal component n;
c) the decomposed low-resonance signal component y of each lead of the electrocardiosignal yLRespectively sending the signals into a neural network for training, and outputting a detection result after the training is finished;
d) and combining the detection results output by the leads, calculating the final classification result by adopting a Voting algorithm, and giving different weights to the leads to obtain the final detection result.
Further, step a) is processed by the following steps:
a-1) slicing a static twelve-lead electrocardiosignal y by using a computer, and intercepting N sampling points in front of each lead to obtain a sliced electrocardiosignal;
a-2) processing the sliced electrocardio data by utilizing a 0.5-40Hz band-pass filter to obtain an electrocardio signal y' with partial noise removed.
Further, step b) is processed by the following steps:
b-1) according to the formula y ═ yH+yL+ n decomposing the partial noise eliminated electrocardiosignal y' based on the resonance sparse decomposition algorithm, wherein yHIs a high resonance signal component, yLIs a low resonance signal component, n is a residual signal component;
b-2) by the formula y ═ W1M1+W2M2+ n calculating partial noise eliminated electrocardiosignal y' with W1Is a high resonance signal component yHIn redundant dictionary M1Transform coefficient of down, W2Is a low resonance signal component yLIn redundant dictionary M2A lower transform coefficient, n being a residual signal component;
b-3) by the formula J (W)1,W2)＝||y′-W1M1-W2M2||2+λ1||W1||1+λ2||W2||2Calculating an objective function J (W) for measuring sparsity1,W2) In the formula of1And λ2As a weight coefficient, M1Is a high resonance signal component yHRedundant dictionary of, M2Is a low resonance signal component yLRedundant dictionaries of (1);
b-4) applying an iterative algorithm to the target function J (W)1,W2) Iteration is carried out to obtain an objective function J (W)1,W2) Using the objective function J (W)1,W2) The minimum value of (A) is obtained as a transformation coefficient W of the high-resonance signal component at that time1Transformation coefficient W of low resonance signal component2By the formula yH＝W1*M1Will result in a high resonance signal component yHIn redundant dictionary M1Is represented by a new coefficient of variation, represented by the formula yL＝W2*M2Will result in a low resonance signal component yLIn redundant dictionary M2Represented by the new coefficient of variation.
Further, step c) is processed by the following steps:
the decomposed low resonance component y of the electrocardiosignalLPerforming network training as input of a neural network model, wherein the neural network model consists of four volume blocks, a bidirectional long-time and short-time memory network and a full connection layer, the bidirectional long-time and short-time memory network is connected behind the four volume blocks, the output dimensionality of the network is set to be 32, the network is converted into a characteristic value x with the output dimensionality of 2 through the full connection layer, and the probability P that a signal is normal is predicted through the characteristic value xNiAnd probability of atrial fibrillation PAFiWherein i is the serial number of different leads of the electrocardiogram data.
Further, step d) is processed by the following steps:
the probability of outputting twelve leads is according to a formulaAndcalculating to obtain final detection result, wherein P isNi+PAFi＝1，ωNiThe normal probability weight, ω, corresponding to the ith leadAFiThe probability weight of atrial fibrillation corresponding to the ith lead.
Preferably, in step a-1), N is 4096.
Preferably, each convolution block of the four convolution blocks includes one convolution layer and one pooling layer, the convolution layer is a one-dimensional convolution layer, the size of the convolution kernel is set to 1 × 31, the step size is set to 1, the padding is set to 15, the dimensionalities of the convolution layers in the four convolution blocks are 8/64/64/8 respectively, the kernel size in the pooling layer is set to 2, and the step size is set to 2.
The invention has the beneficial effects that: according to the characteristics of atrial fibrillation signals, the electrocardiosignals are processed by resonance sparse decomposition, the decomposed low-resonance component parts are sent to a neural network for training, the probabilities obtained by training models are subjected to joint operation through a Voting algorithm (Voting) after lead-by-lead training one by one to obtain a final detection result, other features do not need to be extracted manually additionally when the atrial fibrillation detection is carried out, the network structure is simple, the operation time is shortened, and real-time atrial fibrillation signal detection can be realized. Aiming at the performance characteristics of 'P wave disappears and F wave appears' in the atrial fibrillation signal, the characteristics of the atrial fibrillation signal are highlighted on the basis of resonance sparse decomposition, and the operation time is reduced on the basis of accuracy guarantee through a simple neural network structure.
Drawings
FIG. 1 is a flow chart of a method of the present invention;
FIG. 2 is a graph comparing atrial fibrillation signals before and after resonance sparse decomposition according to the present invention;
FIG. 3 is a comparison of normal signals before and after resonance sparse decomposition according to the present invention;
FIG. 4 is a diagram of a neural network architecture of the present invention;
FIG. 5 is a flow chart of the Voting algorithm of the present invention.
Detailed Description
The invention will be further described with reference to fig. 1 to 4.
A twelve-lead electrocardiosignal atrial fibrillation detection method based on resonance sparse decomposition comprises the following steps:
a) slicing and drying the static twelve-lead electrocardiosignal y by using a computer to obtain an electrocardiosignal y' with fixed length and part of noise removed;
b) processing each lead of the electrocardiosignal y' with partial noise removed by a resonance sparse decomposition technology one by one to obtain a high-resonance signal component yHLow resonance signal component yLAnd a residual signal component n;
c) the decomposed low-resonance signal component y of each lead of the electrocardiosignal yLRespectively sending the signals into a neural network for training, and outputting a detection result after the training is finished;
d) and combining the detection results output by the leads, calculating the final classification result by adopting a Voting algorithm, and giving different weights to the leads to obtain the final detection result.
According to the characteristics of atrial fibrillation signals, the electrocardiosignals are processed by resonance sparse decomposition, the decomposed low-resonance component parts are sent to a neural network for training, the probabilities obtained by training models are subjected to joint operation through a Voting algorithm (Voting) after lead-by-lead training one by one to obtain a final detection result, other features do not need to be extracted manually additionally when the atrial fibrillation detection is carried out, the network structure is simple, the operation time is shortened, and real-time atrial fibrillation signal detection can be realized. Aiming at the performance characteristics of 'P wave disappears and F wave appears' in the atrial fibrillation signal, the characteristics of the atrial fibrillation signal are highlighted on the basis of resonance sparse decomposition, and the operation time is reduced on the basis of accuracy guarantee through a simple neural network structure.
In fig. 2, the abscissa is time and the ordinate is signal amplitude. The truncated signal has a length of 4096 samples and a sampling rate of 500Hz, and therefore has a time length of 8.192 s. Three parts of a high resonance component signal, a low resonance component signal and a residual component signal are obtained through resonance sparse decomposition. The high-resonance component signal mainly comprises a part which is continuously oscillated in the original signal, namely a continuously existing myoelectric noise signal; the low resonance component signal mainly comprises a transient impact part in the original signal, namely an electrocardiosignal appearing periodically; the residual component signal is the remaining portion of the original signal after the removal of the high-resonance component signal and the low-resonance component signal. It can be seen from fig. 2 (atrial fibrillation) that the low resonance component signal significantly reduces much of the glitch-like noise compared to the original signal. And it can be obviously observed that the P wave in the low resonance component signal disappears, and atrial fibrillation occurs. In fig. 3 for normal signal decomposition, it can be seen that the low-resonance component signal contains a complete electrocardiosignal, and the characteristic is prominent, which is helpful for learning of the later neural network.
Step a) is processed by the following steps:
a-1) slicing a static twelve-lead electrocardiosignal y by using a computer, and intercepting N sampling points in front of each lead to obtain a sliced electrocardiosignal;
a-2) processing the sliced electrocardio data by utilizing a 0.5-40Hz band-pass filter to obtain an electrocardio signal y' with partial noise removed.
Step b) is processed by the following steps:
b-1) according to the formula y ═ yH+yL+ n decomposing the partial noise eliminated electrocardiosignal y' based on the resonance sparse decomposition algorithm, wherein yHIs a high resonance signal component, yLIs a low resonance signal component, n is a residual signal component;
b-2) by the formula y ═ W1M1+W2M2+ n calculating partial noise eliminated electrocardiosignal y' with W1Is a high resonance signal component yHIn redundant dictionary M1Transform coefficient of down, W2Is a low resonance signal component yLIn redundant dictionary M2A lower transform coefficient, n being a residual signal component;
b-3) by the formula J (W)1,W2)＝||y′-W1M1-W2M2||2+λ1||W1||1+λ2||W2||2Calculating an objective function J (W) for measuring sparsity1,W2) In the formula of1And λ2As a weight coefficient, M1Is a high resonance signal component yHRedundant dictionary of, M2Is a low resonance signal component yLRedundant dictionaries of (1);
b-4) applying an iterative algorithm to the target function J (W)1,W2) Iteration is carried out to obtain an objective function J (W)1,W2) Using the objective function J (W)1,W2) The minimum value of (A) is obtained as a transformation coefficient W of the high-resonance signal component at that time1Transformation coefficient W of low resonance signal component2By the formula yH＝W1*M1Will result in a high resonance signal component yHIn redundant dictionary M1Is represented by a new coefficient of variation, represented by the formula yL＝W2*M2Will result in a low resonance signal component yLIn redundant dictionary M2Represented by the new coefficient of variation.
Step c) is carried out by the following steps:
the decomposed low resonance component y of the electrocardiosignalLPerforming network training as input of a neural network model, wherein the neural network model consists of four volume blocks, a bidirectional long-time and short-time memory network and a full connection layer, the bidirectional long-time and short-time memory network is connected behind the four volume blocks, the output dimensionality of the network is set to be 32, the network is converted into a characteristic value x with the output dimensionality of 2 through the full connection layer, and the probability P that a signal is normal is predicted through the characteristic value xNiAnd probability of atrial fibrillation PAFiWherein i is the serial number of different leads of the electrocardiogram data.
Step d) is carried out by the following steps:
the probability of outputting twelve leads is according to a formulaAndcalculating to obtain final detection result, wherein P isNi+PAFi＝1，ωNiThe normal probability weight, ω, corresponding to the ith leadAFiThe probability weight of atrial fibrillation corresponding to the ith lead.
The value of N in the step a-1) is 4096.
Each convolution block of the four convolution blocks comprises a convolution layer and a pooling layer, the convolution layer is a one-dimensional convolution layer, the size of a convolution kernel is set to be 1 x 31, the step size is set to be 1, the padding is set to be 15, the dimensionalities of the convolution layers in the four convolution blocks are 8/64/64/8 respectively, the size of the kernel in the pooling layer is set to be 2, and the step size is set to be 2.
- 上一篇：一种医用注射器针头装配设备
- 下一篇：一种基于图卷积网络的脑电图信号分析方法