Implementation method of electrocardiogram atrial fibrillation detection based on signal decomposition and convolution network

Abstract

The implementation method of electrocardiogram atrial fibrillation detection based on signal decomposition and convolution network can form pseudo QRS wave group signals and pseudo T and P wave signals from single lead one-dimensional electrocardiogram data through signal decomposition, can make the field knowledge in the clinical diagnosis of electrocardiogram atrial fibrillation into the feature representation of electrocardiogram, and can make the machine mine the characteristics of R-R interval and P wave form in the electrocardiogram atrial fibrillation through the method of deep convolution network, so that the electrocardiogram atrial fibrillation diagnosis is finally realized. When the method is used for atrial fibrillation detection, the characteristics of the electrocardiogram atrial fibrillation are explicitly provided as input to the neural network, the task of atrial fibrillation detection is divided into two subtasks of investigating the characteristics of the electrocardiogram atrial fibrillation, and the accuracy of the atrial fibrillation detection is greatly improved, wherein the accuracy of the detection method of the deep convolution network using the ordinary convolution kernel is 0.836, and the average F1 score is 0.813; the accuracy of the detection method of the deep convolution network using the hollow convolution kernel is 0.867, and the average F1 score is 0.843.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of automatic detection of atrial fibrillation, and relates to an implementation of atrial fibrillation signal detection based on single-lead electrocardiogram, in particular to an implementation method of single-lead electrocardiogram atrial fibrillation detection based on a deep learning method. BACKGROUND

[0002] The electrocardiogram of a patient with atrial fibrillation is usually characterized by the disappearance of P waves, the appearance of f waves, unequal R-R intervals and narrow QRS waves, but in clinical practice, atrial fibrillation with f wave disappearance or P wave non-disappearance can also be observed. In recent years, with the development of wearable smart devices, the measurement of single-lead electrocardiogram has become increasingly convenient and has gradually entered the daily use of individuals, families and sports. However, due to the lack of professional knowledge, the effective utilization of these data has become a problem. With the development of machine learning and deep learning, people have begun to study automatic algorithms for atrial fibrillation detection. Traditional atrial fibrillation detection algorithms often choose one of two features, such as identifying the position of R peaks, calculating the length of R-R intervals and identifying their regularity using certain machine learning algorithms, or constructing a QRS complex template to eliminate ventricular activity in order to better identify and classify atrial activity, especially P wave morphology. Such methods rely only on one of the two features and are easily affected by the instability of a single feature.

[0003] In recent years, deep learning-based atrial fibrillation detection methods have emerged. Deep learning has been quite mature in speech, image and text processing and has achieved many results. Applying deep learning to electrocardiogram atrial fibrillation detection reduces the burden of feature design to some extent and can extract more complex features from electrocardiogram data. On the other hand, this completely discards the human experience accumulated over a long period of clinical electrocardiogram diagnosis. The training of neural networks in deep learning has considerable randomness, and the differences produced in initialization can cause the network to learn features that are not universal or even incorrect, resulting in a decline in network performance. Such problems leave room for improvement in the accuracy of existing deep learning-based electrocardiogram atrial fibrillation detection algorithms. SUMMARY

[0004] The purpose of the present application is to solve the above problems existing in the background art, provide an implementation method of single-lead electrocardiogram atrial fibrillation detection based on signal decomposition and convolution network, which can convert one-dimensional electrocardiogram data into two signals, namely pseudo QRS complex signal and pseudo T, P wave signal, through signal decomposition. Based on such signal conversion, the deep features of electrocardiogram are learned using a hollow convolution residual network, and the corresponding classification of the signal is performed. When the method of the present application is used for atrial fibrillation detection of electrocardiogram, the professional knowledge in the field of clinical electrocardiogram atrial fibrillation diagnosis is utilized, the characteristics of atrial fibrillation electrocardiogram are utilized more completely, and the accuracy of single-lead electrocardiogram atrial fibrillation detection is greatly improved, which has important role and significance for the development of wearable device technology and personal health.

[0005] The purpose of the present application is realized by the following technical solutions:

[0006] The implementation method of electrocardiogram atrial fibrillation detection based on signal decomposition and convolution network, the method comprises:

[0007] (1) using a single-lead electrocardiogram (ECG) acquisition device to collect single-lead electrocardiogram data of multiple users;

[0008] (2) segmenting the electrocardiogram data of each user, the number of sampling points of each segment of single-lead electrocardiogram data is n, and each segment of single-lead electrocardiogram data is taken as a unit for atrial fibrillation detection, wherein n is adjusted according to a specific application scenario;

[0009] (3) preprocessing each segment of single-lead electrocardiogram data to remove noise of the signal;

[0010] (4) performing signal inversion detection on the single-lead electrocardiogram data to correct abnormal signals generated by electrode inversion;

[0011] (5) determining the position of the R peak in the electrocardiogram segment by means of artificial labeling or automatic algorithm, and calculating the R-R interval length;

[0012] (6) performing signal segment decomposition operation on the electrocardiogram signal based on the R peak position and the R-R interval length to form a pseudo QRS complex signal segment and a pseudo T, P wave signal segment with n sample points;

[0013] (7) repeating steps (2)-(6) for each segment of single-lead electrocardiogram obtained from a user to convert the continuously collected electrocardiogram data into a pseudo QRS complex signal segment and a pseudo T, P wave signal segment with n sample points as the input of the deep convolution network;

[0014] (8) Repeat steps (1)-(7) several times for each user's electrocardiogram data to obtain the signal decomposition results of the electrocardiogram data segment. Use this to train the deep convolutional neural network, obtain the final parameters, and form a deep convolutional network model for atrial fibrillation detection.

[0015] Furthermore, in step (2), after segmentation, if the number of sample points in the last segment is less than n, and it is less than n / 2, then it is discarded; otherwise, the signal segment is looped twice to make the total number of sample points greater than n, and the first n sample points are taken as the last segment of the electrocardiogram signal.

[0016] Furthermore, in step (4), the signal reversal detection method is based on R-peak positioning and the determination of the extreme values ​​during the RR cycle to determine the signal reversal.

[0017] Furthermore, in step (6), the specific process of decomposing the single-lead electrocardiogram signal segment is as follows:

[0018] a. For each ECG signal x[n], determine the position of the R-peak in the signal using manual annotation or an automated algorithm, and denote the set of R-peak sample points as r = {r1, r2, ..., r...} k}, where 0 < r1 < r2 < ... < r k <n, k is the number of sample points for the R-peak position determined by manual annotation or automated algorithm;

[0019] b. Let i = 1, 2, ..., k-1, and let l i =r i+1 -r i , l i That is, the length of the i-th RR interval in the electrocardiogram signal, and denoted as l = {l1, l2, ..., l...} k-1};

[0020] c. Take the 5th percentile of l, denoted as

[0021] d. Let i = 1, 2, ..., k-1, and let make [p i ,q i That is, the range of the i-th QRS complex in the electrocardiogram signal segment;

[0022] e. Record make make x obtained at this time QRS [n] is the pseudo-QRS group signal, x TP [n] refers to the pseudo-T and P wave signals.

[0023] Furthermore, in step (8), the deep convolutional neural network is built based on Tensorflow.

[0024] The advantages of this invention over the prior art are as follows:

[0025] 1. This invention introduces domain knowledge from clinical electrocardiogram diagnosis into the traditional atrial fibrillation detection neural network. By determining the position of the R peak, different stages in the electrocardiogram cycle are divided to form pseudo-QRS complex signal segments and pseudo-T and P wave signal segments. This data processing decomposes the atrial fibrillation detection task into the identification task of RR interval features and P wave morphology features in the electrocardiogram, reducing the difficulty of network learning. Compared with ordinary convolutional residual networks, it has an improvement of about 5% in F1 score.

[0026] 2. This invention was tested on the training dataset of the PhysioNet / CinC 2017 challenge and achieved good performance. Specifically, the detection method using a deep convolutional network with ordinary convolutional kernels achieved an accuracy of 0.836 and an average F1 score of 0.813; the detection method using a deep convolutional network with dilated convolutional kernels achieved an accuracy of 0.867 and an average F1 score of 0.843. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating the overall implementation method of the electrocardiogram atrial fibrillation detection based on signal decomposition and convolutional networks using clinical data, as described in this invention.

[0028] Figure 2 This is a flowchart illustrating the overall implementation method of the present invention for detecting atrial fibrillation on electrocardiogram using existing datasets based on signal decomposition and convolutional networks.

[0029] Figure 3 This is a comparison diagram of the input and output of the signal decomposition method described in this invention;

[0030] Figure 4 This is a network structure diagram of the deep convolutional neural network for atrial fibrillation detection using ordinary convolutional kernels as described in this invention;

[0031] Figure 5 This is a network structure diagram of the deep convolutional neural network for atrial fibrillation detection using dilated convolutional kernels as described in this invention. Detailed Implementation

[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments, but it is not limited thereto. Any modifications or equivalent substitutions to the technical solution of the present invention that do not depart from the spirit and scope of the technical solution of the present invention should be covered within the protection scope of the present invention.

[0033] The method of the present application can form pseudo QRS wave group signals and pseudo T and P wave signals from single-lead one-dimensional electrocardio data through signal decomposition, enable domain knowledge in the field of clinical diagnosis of electrocardiogram atrial fibrillation to be incorporated into the feature representation of electrocardiogram, and enable a machine to mine the characteristics of R-R interval and P wave form in atrial fibrillation electrocardiogram through a deep convolution network method, so as to finally realize electrocardiogram atrial fibrillation diagnosis. When the method of the present application is used for atrial fibrillation detection, the characteristics of atrial fibrillation electrocardiogram are explicitly provided as input to a neural network, the task of atrial fibrillation detection is decomposed into subtasks of examining two atrial fibrillation electrocardiogram characteristics, and the accuracy of atrial fibrillation detection is greatly improved, in which the accuracy of the detection method of the deep convolution network using a general convolution kernel is 0.836, and the average F1 score is 0.813; the accuracy of the detection method of the deep convolution network using a hollow convolution kernel is 0.867, and the average F1 score is 0.843.

[0034] Embodiment 1

[0035] As shown in the following Figure 1 , the present embodiment provides an implementation method of atrial fibrillation detection based on signal decomposition and deep convolution neural network using clinical data. In order to introduce the domain knowledge of clinical atrial fibrillation detection, the task of atrial fibrillation detection is decomposed into subtasks of detecting two characteristics of atrial fibrillation electrocardiogram, the difficulty of extracting electrocardiogram characteristics by the network is reduced, and the accuracy of detection is improved. The deep convolution neural network provided in the present embodiment is constructed based on one of the deep learning frameworks widely used in academia and industry, namely Tensorflow. The specific implementation steps are as follows:

[0036] (1) The single-lead electrocardiogram data of the user is measured by a wearable electrocardiogram acquisition device, and the signal is segmented with a certain number of sample points n, so that the number of sample points of each segment of electrocardiogram signal is n. For the last signal, if its length is less than n / 2, it is discarded; otherwise, the signal segment is spliced with itself, and the first n sample points are taken as the last segment of electrocardiogram signal.

[0037] (2) Each segment of electrocardiogram data is classified by a medical practitioner with relevant experience, and the signal is divided into four categories: normal signal, atrial fibrillation signal, other disease signal, and noise signal.

[0038] (3) Signal inversion detection is performed on each segment of electrocardiogram signal, and the specific operation is as follows:

[0039] a. For each electrocardiogram chamber signal segment x[n], the position of the R peak in the signal is determined using artificial labeling or automatic algorithm, and the R peak sample point set is denoted as p = {p1, p2, …, p k}, where 0 < p1 < p2 < … < p k<n, k is the number of sample points for the R-peak position determined by manual annotation or automated algorithm;

[0040] b. Let i = 1, 2, ..., k-1, for each (p i ,p i+1 Given an interval, find the point b with the smallest value within that interval. i ,Right now And let b = {b1, b2, ..., b} k-1};

[0041] c. Let i = 1, 2, ..., k-1, and let s i =p i +b1, which is the sum of the values ​​of the nearest peaks and troughs, and denoted as s = {s1, s2, ..., s}. k-1};

[0042] d. Let the average value of s be... like If s is less than zero, it is considered that the signal x[n] has experienced electrode reversal during the measurement process, and the signal is reversed, i.e., x'[i] = 0 - x[i], i = 1, 2, ..., n, where x'[n] is the corrected signal; if s is greater than or equal to zero, the signal measurement is considered to be correct, and no operation is performed.

[0043] (4) Figure 3 As shown, signal decomposition is performed on any ECG data segment obtained in step (3). The specific operation is as follows:

[0044] a. For each ECG signal segment x[n], determine the position of the R-peak in the signal using manual annotation or an automated algorithm, and denote its sample point set as r = {r1, r2, ..., r...} k}, where 0 <r1<r2<…<r k <n, k is the number of sample points for the R-peak position determined by manual annotation or automated algorithm;

[0045] b. Let i = 1, 2, ..., k-1, and let l i =r i+1 -r i , l i That is, the length of the i-th RR interval in the electrocardiogram signal segment, and denoted as l = {l1, l2, ..., l...} k-1};

[0046] c. Take the 5th percentile of l, denoted as

[0047] d. Let i = 1, 2, ..., k-1, and let make [p i ,qi That is, the range of the i-th QRS complex in the electrocardiogram signal segment;

[0048] e. Record make make x obtained at this time QRS [n] is the pseudo-QRS group signal, x TP [n] refers to the pseudo-T and P wave signals.

[0049] (5) The pseudo QRS complex signal and pseudo T and P wave signals generated by the signal decomposition of each ECG signal segment are spliced ​​together to form a two-dimensional signal of size n×2 as the input of the neural network.

[0050] (6) Repeat steps (3)-(5) to convert a continuous ECG data segment into a neural network system data input with multiple sample points of n, which is composed of pseudo-QRS wave group signals and pseudo-T and P wave signals, and carries one of four labels.

[0051] (7) Repeat steps (1)-(6) to obtain the ECG data segment used to train the deep convolutional neural network.

[0052] (8) Divide the dataset into three parts: 75% of the total number of ECG signal segments as the training set, 15% as the validation set, and 10% as the test set.

[0053] (9) Construct and train the neural network, where:

[0054] like Figure 4 As shown, the structure of the neural network is as follows:

[0055] a. First, the input signal passes through a one-dimensional convolutional layer. The stride of the convolutional layer is 2, the kernel size is 5, and the number of kernels is 32. Then, the signal passes through a batch regularization layer, and finally through a leaky modified linear unit activation function with a parameter of 0.1.

[0056] b. The signal passes through a residual block. The residual block has two paths, one of which is left untouched. On the other path, the signal passes through three ordinary one-dimensional convolutional layers with a stride of 1 and a kernel size of 5, the number of kernels equal to the number of input channels. Before each convolutional layer, the signal passes through a batch regularization layer. After each convolutional layer, the signal passes through a modified linear unit activation function with leakage and a parameter of 0.1. The input is summed after passing through both paths, and the sum is the output.

[0057] c. Repeat steps a and b 7 times, but adjust the number of convolution kernels in the convolutional layer in a. When performing operation a from the 2nd to the 8th time, the number of convolution kernels is 64, 96, 128, 256, 512, 256, and 128, respectively.

[0058] d. The input signal passes through a one-dimensional convolutional layer. The stride of the convolutional layer is 2, the kernel size is 5, and the number of kernels is 96. The signal then passes through a batch regularization layer, and finally through a leaky modified linear unit activation function with a parameter of 0.1.

[0059] e. The signal passes through a one-dimensional convolutional layer. The stride of the convolutional layer is 1, the kernel size is 1, and the number of kernels is 4. Subsequently, the signal is subjected to temporal averaging, that is, the signal is averaged in the first dimension.

[0060] f. The signal passes through a SoftMax activation layer, outputting the final prediction result.

[0061] Based on this, all weights in the model are initialized using the standard Glorot initializer, and all biases in the network are zero-initialized. Simultaneously, to avoid or mitigate overfitting, random deactivation with a parameter of 0.2 is used for each layer of the network. The network is trained using the Adam optimizer, with an initial learning rate of 0.0001, β1 of 0.9, β2 of 0.999, and e of 1e⁻⁷. The number of iterations for network training is determined using descent and early stopping. If the loss function does not decrease within 4 iterations, the learning rate is reduced to 10% of its original value; if it does not decrease within 12 iterations, network training is terminated.

[0062] (10) In actual atrial fibrillation detection applications, steps (1)-(7) are performed on each patient to obtain pseudo-QRS complex signals and pseudo-T and P wave signals after decomposition of the electrocardiogram signal, without the need to judge the signal category. Then, the deep convolutional neural network parameter model trained in step (8) is used to detect the signal, so that atrial fibrillation can be determined for a continuous single-lead electrocardiogram signal. On a complete signal, as long as a segment of electrocardiogram signal with n sample points is classified as an atrial fibrillation signal, atrial fibrillation can be determined for the user's electrocardiogram data.

[0063] Example 2:

[0064] like Figure 2 As shown, this embodiment provides an implementation method for atrial fibrillation detection based on signal decomposition and a deep dilated convolutional neural network using the PhysioNet / CinC 2017 challenge dataset. This embodiment is built based on Tensorflow. The specific implementation steps are as follows:

[0065] (1) Let the size of the dataset be m, and let i = 1, 2, ..., m. Let the number of sample points of the i-th signal be n. i .make That is, the length of the longest signal in the dataset.

[0066] (2) Let i = 1, 2, ..., m. If the length of the i-th signal is n... i If the value is less than n, then the signal segment is spliced ​​with itself. Then, the first n sample points are taken as the i-th electrocardiogram signal.

[0067] (3) Perform signal inversion detection on each electrocardiogram signal. The specific operation is as follows:

[0068] a. For each ECG signal x[n], determine the position of the R-peak in the signal using manual annotation or an automated algorithm, and denote its sample point set as p = {p1, p2, ..., pn}. k}, where 0 <p1<p2<…<p k <n, k is the number of sample points for the R-peak position determined by manual annotation or automated algorithm;

[0069] b. Let i = 1, 2, ..., k-1, for each (p i ,p i+1 Given an interval, find the point b with the smallest value within that interval. i ,Right now And let b = {b1, b2, ..., b} k-1};

[0070] c. Let i = 1, 2, ..., k-1, and let s i =p i +b1, which is the sum of the values ​​of the nearest peaks and troughs, and denoted as s = {s1, s2, ..., s}. k-1};

[0071] d. Let the average value of s be... like If the value is less than zero, it is considered that the signal x[n] has experienced electrode reversal during the measurement process, and the signal is inverted, i.e., x'[i] = 0 - x[i], i = 1, 2, ..., n, where x'[n] is the corrected signal; if If the value is greater than or equal to zero, the signal measurement is considered to be correct, and no further action is required.

[0072] (4) Figure 3 As shown, signal decomposition is performed on any ECG data obtained in step (3). The specific operation is as follows:

[0073] a. For each ECG signal segment x[n], determine the position of the R-peak in the signal using manual annotation or an automated algorithm, and denote its sample point set as r = {r1, r2, ..., r...} k}, where 0 <r1<r2<…<r k <n, k is the number of sample points for the R-peak position determined by manual annotation or automated algorithm;

[0074] b. Let i = 1, 2, ..., k-1, and let l i =r i+1 -r i , l i That is, the length of the i-th RR interval in the electrocardiogram signal segment, and denoted as l = {l1, l2, ..., l...} k-1};

[0075] c. Take the 5th percentile of l, denoted as

[0076] d. Let i = 1, 2, ..., k-1, and let make [p i ,q i That is, the range of the i-th QRS complex in the electrocardiogram signal segment;

[0077] e. Record make make x obtained at this time QRS [n] is the pseudo-QRS group signal, x TP [n] refers to the pseudo-T and P wave signals.

[0078] (5) The pseudo QRS complex signal and pseudo T and P wave signals generated by the signal decomposition of each ECG signal segment are spliced ​​together to form a two-dimensional signal of size n×2 as the input of the neural network.

[0079] (6) Repeat steps (3)-(5) several times for each single-lead electrocardiogram obtained from a user, and convert all the electrocardiogram data in the dataset into a neural network system data input with n sample points, which is composed of pseudo-QRS wave group signals and pseudo-T and P wave signals, and has one of four labels.

[0080] (7) Repeat steps (1)-(6) several times for each user's electrocardiogram data to obtain a dataset of electrocardiogram data segments for training the deep convolutional neural network.

[0081] (8) Divide the dataset into three parts: 75% of the total number of ECG signal segments as the training set, 15% as the validation set, and 10% as the test set.

[0082] (9) Construct a neural network and train it, wherein:

[0083] As shown in Figure 5 , the structure of the neural network is as follows:

[0084] a. First, the input signal passes through a one-dimensional convolutional layer. The step size of the convolutional layer is 2, the convolution kernel size is 5, and the number of convolution kernels is 32. Then the signal passes through a batch normalization layer, and finally passes through a leaky rectified linear unit activation function with a parameter of 0.1;

[0085] b. The signal passes through a residual block. There are two paths in the residual block, one of which does not operate. In the other path, the signal passes through three hollow one-dimensional convolutional layers, the step size of the convolutional layer is 1, the convolution kernel size is 5, and the number of convolution kernels is the same as the number of input channels. The expansion rates of the three hollow convolutional layers are 1, 2, and 4, respectively. Before each convolutional layer, the signal first passes through a batch normalization layer, and after each convolutional layer, the signal passes through a leaky rectified linear unit activation function with a parameter of 0.1. The sum of the inputs after passing through the two paths is taken as the output;

[0086] c. The operations of steps a-b are repeated 7 times, but the number of convolution kernels in the convolutional layer in a is adjusted. The number of convolution kernels in the second to eighth times of operation a is 64, 96, 128, 256, 512, 256, and 128, respectively;

[0087] d. The input signal passes through a one-dimensional convolutional layer. The step size of the convolutional layer is 2, the convolution kernel size is 5, and the number of convolution kernels is 96. Then the signal passes through a batch normalization layer, and finally passes through a leaky rectified linear unit activation function with a parameter of 0.1;

[0088] e. The signal passes through a one-dimensional convolutional layer. The step size of the convolutional layer is 1, the convolution kernel size is 1, and the number of convolution kernels is 4. Then, the signal is processed by time average, that is, the average value of the signal in the first dimension is taken;

[0089] f. The signal passes through a SoftMax activation layer to output the final prediction result.

[0090] On this basis, all weights in the model are initialized using the standard Glorot initializer, and all biases in the network are initialized to zero. At the same time, in order to avoid or alleviate the phenomenon of overfitting, for each layer of the network, random inactivation with a parameter of 0.2 is used. The network is trained using the Adam optimizer, with an initial learning rate of 0.0001, β1 of 0.9, β2 of 0.999, and e of 1e-07. The flat descent and early stopping method are used to determine the number of iterations of network training, and if the loss function does not decrease in 4 iterations, the learning rate will be reduced to 10% of the original, and if it does not decrease in 12 iterations, the training of the network will be ended.

[0091] (10) In actual atrial fibrillation detection applications, the single-lead electrocardiogram data of the user is measured by a wearable electrocardiogram acquisition device, the signal is segmented by the number of sample points n, so that the number of sample points of each segment of electrocardiogram signal is not greater than n. Then, the operations of steps (1)-(7) are performed to obtain the pseudo QRS complex signal and the pseudo T and P wave signal after decomposition of the electrocardiogram signal, without the need for signal category judgment. Then, the deep convolutional neural network parameter model trained in step (9) is used to detect the signal, so that a continuous single-lead electrocardiogram signal can be determined for atrial fibrillation. As long as a segment of electrocardiogram signal with a sample point number of n is classified as an atrial fibrillation signal on a complete signal, the electrocardiogram data of the user is determined for atrial fibrillation.

Claims

1. A method for detecting atrial fibrillation on electrocardiogram based on signal decomposition and convolutional networks, characterized in that: The method is as follows: (1) Collect single-lead electrocardiogram (ECG) data from multiple users using a single-lead ECG acquisition device; (2) The ECG data of each user is segmented, and the number of sampling points of each single-lead ECG data segment is n. Each single-lead ECG data segment is used as the unit for atrial fibrillation detection. (3) Preprocess each single-lead ECG data segment to remove signal noise; (4) Perform signal inversion detection on single-lead electrocardiogram data and correct abnormal signals generated by electrode reversal; (5) Determine the position of the R peak in the electrocardiogram segment by manual annotation or automated algorithm, and calculate the length of the RR interval; (6) Based on the position of the R peak and the length of the RR interval, the ECG signal is decomposed into signal segments to form pseudo-QRS complex signal segments and pseudo-T and P wave signal segments with n sample points; (7) Repeat steps (2)-(6) several times for each single-lead electrocardiogram obtained from a user, and convert the continuously acquired electrocardiogram data into pseudo QRS complex signal segments and pseudo T and P wave signal segments with n sample points, which are used as inputs to the deep convolutional network. (8) Repeat steps (1)-(7) several times for each user's electrocardiogram data to obtain the signal decomposition results of the electrocardiogram data segment. Use this to train the deep convolutional neural network, obtain the final parameters, and form a deep convolutional network model for atrial fibrillation detection.

2. The method for detecting atrial fibrillation on electrocardiogram based on signal decomposition and convolutional networks according to claim 1, characterized in that: In step (2), after segmentation, if the number of sample points in the last segment is less than n, and it is less than n / 2, then it is discarded; otherwise, the signal segment is looped twice to make the total number of sample points greater than n, and the first n sample points are taken as the last segment of the electrocardiogram signal.

3. The method for detecting atrial fibrillation on electrocardiogram based on signal decomposition and convolutional networks according to claim 1, characterized in that: In step (4), the signal reversal detection method is based on R-peak positioning and the determination of the extreme values ​​during the RR cycle to determine the reversal of the signal.

4. The method for detecting atrial fibrillation on electrocardiogram based on signal decomposition and convolutional networks according to claim 1, characterized in that: In step (6), the specific process of decomposing the single-lead electrocardiogram signal segment is as follows: a. For each ECG signal x[n], determine the position of the R-peak in the signal using manual annotation or an automated algorithm, and denote the set of R-peak sample points as r = {r1, r2, ..., r...} k }, where 0 < r1 < r2 < ... < r k <n, k is the number of sample points for the R-peak position determined by manual annotation or automated algorithm; b. Let i = 1, 2, ..., k-1, and let l i =r i+1 -r i , l i That is, the length of the i-th RR interval in the electrocardiogram signal, and denoted as l = {l1, l2, ..., l...} k-1 }; c. Take the 5th percentile of l, denoted as d. Let i = 1, 2, ..., k-1, and let make [p i ,q i That is, the range of the i-th QRS complex in the electrocardiogram signal segment; e. Record make make x obtained at this time QRS [n] is the pseudo-QRS group signal, x TP [n] refers to the pseudo-T and P wave signals.

5. The method for detecting atrial fibrillation on electrocardiogram based on signal decomposition and convolutional networks according to claim 1, characterized in that: In step (8), the deep convolutional neural network is built based on Tensorflow.

Images