CN109171712B

CN109171712B - Atrial fibrillation identification method, atrial fibrillation identification device, atrial fibrillation identification equipment and computer readable storage medium

Info

Publication number: CN109171712B
Application number: CN201811138645.6A
Authority: CN
Inventors: 李鸣春; 孙亮; 何光宇
Original assignee: Neusoft Corp
Current assignee: Neusoft Corp
Priority date: 2018-09-28
Filing date: 2018-09-28
Publication date: 2022-03-08
Anticipated expiration: 2038-09-28
Also published as: CN109171712A

Abstract

The invention provides an atrial fibrillation identification method, an atrial fibrillation identification device, atrial fibrillation identification equipment and a computer readable storage medium, wherein the peak position of an R wave in an electrocardiosignal to be identified is obtained, and then heart rate variability characteristics and heart beat segments are determined according to the peak position of the R wave; acquiring heart beat characteristics of each heart beat segment to form a heart beat characteristic group; and finally, inputting the heart rate variability characteristics and the heart beat characteristic group into an atrial fibrillation recognition model to obtain an atrial fibrillation recognition result. According to the atrial fibrillation recognition method, the heart rate variability characteristics and the heart beat characteristics of the electrocardiosignals are input into the pre-trained atrial fibrillation recognition model, so that atrial fibrillation can be effectively recognized, and the reliability and the accuracy of atrial fibrillation recognition are improved.

Description

Atrial fibrillation identification method, atrial fibrillation identification device, atrial fibrillation identification equipment and computer readable storage medium

Technical Field

The invention relates to the technical field of electrocardio monitoring, in particular to an atrial fibrillation identification method, device and equipment and a computer readable storage medium.

Background

Atrial fibrillation (atrial fibrillation for short) is the most common persistent chronic arrhythmia, often resulting from chaotic atrial activity and irregular atrial compression. The incidence of atrial fibrillation increases with age. Atrial fibrillation is closely related to various diseases such as stroke, heart failure, coronary heart disease, thrombus and the like, and early atrial fibrillation identification can help patients to find heart abnormality in time and reduce disability rate and mortality rate caused by heart diseases.

In the prior art, whether atrial fibrillation occurs or not can be identified by analyzing electrocardiosignals, but the current atrial fibrillation identification method based on heart rate variability is difficult to reach the standard of clinical diagnosis in terms of reliability and accuracy. Some atrial fibrillation recognition methods based on traditional machine learning appear in recent years, but because atrial fibrillation signals are complex, only few features extracted manually are combined with the traditional machine learning method, and atrial fibrillation cannot be recognized accurately.

Disclosure of Invention

The invention provides an atrial fibrillation identification method, an atrial fibrillation identification device, atrial fibrillation identification equipment and a computer readable storage medium, which are used for effectively identifying atrial fibrillation and improving the reliability and accuracy of atrial fibrillation identification.

In a first aspect, the present invention provides a method for atrial fibrillation identification, including:

acquiring the peak position of an R wave in the electrocardiosignal to be identified;

determining heart rate variability features and heart beat fragments according to the peak positions of the R waves;

acquiring heart beat characteristics of each heart beat segment to form a heart beat characteristic group;

and inputting the heart rate variability features and the heart beat feature group into an atrial fibrillation recognition model to obtain an atrial fibrillation recognition result.

Further, the determining the heart rate variability feature according to the peak position of the R wave comprises:

obtaining RR intervals according to the peak positions of adjacent R waves;

acquiring a heart rate variability feature from all of the RR intervals.

Further, the determining the heartbeat segment according to the peak position of the R wave comprises:

acquiring a first RR interval between the peak position of any R wave in the electrocardiosignals to be identified and the peak position of the previous adjacent R wave and a second RR interval between the peak position of the next adjacent R wave and the peak position of the next adjacent R wave;

and cutting the rear 1/3 part of the first RR interval and the front 2/3 part of the second RR interval to be used as the heart beat segment corresponding to the R wave.

Further, the acquiring the heartbeat characteristics of each heartbeat segment includes:

filtering high and low frequency noise in the electrocardiosignal of the heartbeat segment;

normalizing the filtered electrocardiosignals;

and inputting the electrocardiosignals subjected to normalization processing into a heart beat feature extraction model to obtain heart beat features.

Further, the method further comprises:

obtaining first cardiac electrical signal samples of different heart beat types, the heart beat types including a normal heart beat, a left bundle branch block, a right bundle branch block, an atrial premature beat, a ventricular premature beat, and a paced heart beat;

preprocessing each of the first cardiac signal samples, wherein the preprocessing comprises: dividing the first electrocardiosignal sample into heart beat fragments, and filtering and normalizing high and low frequency noise;

establishing a network structure of a convolutional neural network, inputting each preprocessed first electrocardiosignal sample into the convolutional neural network for training, and generating a heartbeat classification model;

and acquiring a heart beat feature extraction model according to the heart beat classification model.

Further, the method further comprises:

obtaining second samples of electrical signals of different rhythm types, wherein the rhythm types include normal rhythm, atrial fibrillation rhythm and other rhythms, and the other rhythms are arrhythmia except for atrial fibrillation rhythm;

acquiring heart rate variability characteristics and a heart beat characteristic group of each second electrocardiosignal sample;

and training the heart rate variability features and the heart beat feature group of the second cardiac signal samples to generate the atrial fibrillation recognition model.

Further, after obtaining a second cardiac signal sample of a different cardiac rhythm type, the method further comprises:

and intercepting a plurality of atrial fibrillation rhythm segments from the second electric signal samples of the atrial fibrillation rhythm in a random sliding mode to serve as second electric signal samples of the new atrial fibrillation rhythm.

Further, the acquiring a heartbeat feature set of each second cardiac signal sample includes:

segmenting heartbeat segments for each second electrocardiosignal sample, and carrying out high-low frequency noise filtering and normalization processing; and inputting the heart beat characteristic into a heart beat characteristic extraction model, and extracting heart beat characteristics to obtain a heart beat characteristic group of the second electrocardiosignal sample.

Further, before the inputting the heart rate variability feature and the heart beat feature set into the atrial fibrillation recognition model, the method further includes:

performing dimensionality reduction processing on the heart beat feature set, and taking a dimensionality reduction processing result as the heart beat feature set;

before training the heart rate variability features and the heart beat feature group of the second cardiac signal sample, the method further comprises:

and performing dimensionality reduction processing on the heartbeat feature group of the second electrocardiosignal sample, and taking a dimensionality reduction processing result as the heartbeat feature group of the second electrocardiosignal sample.

A second aspect of the present invention provides an atrial fibrillation identifying apparatus comprising:

the peak position acquisition module is used for acquiring the peak position of the R wave in the electrocardiosignal to be identified;

the heart rate variability feature acquisition module is used for determining a heart rate variability feature according to the peak position of the R wave;

the heartbeat segment acquisition module is used for determining heartbeat segments according to the peak positions of the R waves;

the heart beat feature acquisition module is used for acquiring heart beat features of each heart beat segment to form a heart beat feature group;

and the atrial fibrillation recognition module is used for inputting the heart rate variability features and the heart beat feature group into an atrial fibrillation recognition model and acquiring an atrial fibrillation recognition result.

Further, the heart rate variability feature acquisition module is configured to:

obtaining RR intervals according to the peak positions of adjacent R waves;

acquiring a heart rate variability feature from all of the RR intervals.

Further, the heartbeat segment acquisition module is configured to:

Further, the heartbeat feature acquisition module is configured to:

normalizing the filtered electrocardiosignals;

Further, the heartbeat feature acquisition module is further configured to:

Further, the atrial fibrillation identification module is further configured to:

before the heart rate variability features and the heart beat feature set are input into an atrial fibrillation recognition model, performing dimensionality reduction on the heart beat feature set, and taking a dimensionality reduction processing result as the heart beat feature set;

before the heart rate variability features and the heart beat feature group of the second electrocardiosignal sample are trained, performing dimensionality reduction processing on the heart beat feature group of the second electrocardiosignal sample, and taking a dimensionality reduction processing result as the heart beat feature group of the second electrocardiosignal sample.

A third aspect of the invention provides an atrial fibrillation identifying device comprising:

a memory;

a processor; and

a computer program;

wherein the computer program is stored in the memory and configured to be executed by the processor to implement the method of the first aspect.

A fourth aspect of the present invention is to provide a computer-readable storage medium having stored thereon a computer program;

which when executed by a processor implements the method according to the first aspect.

According to the atrial fibrillation identification method, the atrial fibrillation identification device, the atrial fibrillation identification equipment and the computer readable storage medium, the peak position of an R wave in an electrocardiosignal to be identified is obtained, and then the heart rate variability feature and the heart beat segment are determined according to the peak position of the R wave; acquiring heart beat characteristics of each heart beat segment to form a heart beat characteristic group; and finally, inputting the heart rate variability characteristics and the heart beat characteristic group into an atrial fibrillation recognition model to obtain an atrial fibrillation recognition result. According to the atrial fibrillation recognition method, the heart rate variability characteristics and the heart beat characteristics of the electrocardiosignals are input into the pre-trained atrial fibrillation recognition model, so that atrial fibrillation can be effectively recognized, and the reliability and the accuracy of atrial fibrillation recognition are improved.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.

FIG. 1 is a flow chart of a method for atrial fibrillation identification according to an embodiment of the present invention;

FIG. 2 is a schematic illustration of heart beat segment segmentation provided by another embodiment of the present invention;

FIG. 3 is a flow chart of a method for atrial fibrillation identification according to another embodiment of the present invention;

FIG. 4 is a flow chart of a method for atrial fibrillation identification according to another embodiment of the present invention;

FIG. 5a is a schematic illustration of a heartbeat segment with a maximum RR interval as provided by another embodiment of the present invention;

FIG. 5b is a schematic diagram of the end-to-end 0 filling of the cardiac segment according to the maximum RR interval according to another embodiment of the present invention;

FIG. 6 is a block diagram of a network structure of a convolutional neural network of a heart beat classification model according to another embodiment of the present invention;

FIG. 7 is a diagram illustrating training results of the heart beat classification model shown in FIG. 6;

FIG. 8 is a schematic diagram of a visualized heart beat classification representation of the heart beat classification model shown in FIG. 6;

FIG. 9 is a flow chart of a method of atrial fibrillation identification according to another embodiment of the present invention;

FIG. 10 is a schematic diagram of data enhancement of atrial fibrillation signals with random sliding according to another embodiment of the present invention;

fig. 11 is a schematic diagram of a heart rhythm feature extraction process according to another embodiment of the present invention;

FIG. 12 is a general framework for a method of atrial fibrillation identification according to another embodiment of the present invention;

FIG. 13 is a block diagram of an atrial fibrillation identification device provided in accordance with an embodiment of the present invention;

fig. 14 is a block diagram of an atrial fibrillation identification device according to another embodiment of the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Fig. 1 is a flow chart of an atrial fibrillation identification method provided by an embodiment of the invention. The embodiment provides an atrial fibrillation identification method, which comprises the following specific steps:

s101, obtaining the peak position of the R wave in the electrocardiosignal to be identified.

The atrial fibrillation recognition method provided by the embodiment can be applied to monitoring equipment such as wearable electrocardio equipment or a medical electrocardiograph, and can be used for collecting electrocardiosignals in real time through a collecting device in the monitoring equipment such as the wearable electrocardio equipment or the medical electrocardiograph, and certainly can also be used for analyzing the electrocardiosignals collected in advance. In this embodiment, the peak position of the R wave in the electrocardiographic signal to be identified can be obtained by a differential threshold method. Wherein, the QRS wave complex (QRS wave complex) of the electrocardiographic signal reflects the changes of the left and right ventricular depolarization potentials and time, wherein the first downward wave is a Q wave, the upward wave is an R wave, and the downward wave is an S wave.

S102, determining heart rate variability features and heart beat fragments according to the peak positions of the R waves.

In this embodiment, by acquiring peak positions of all R waves in the electrocardiographic signal, a series of RR intervals (that is, intervals between adjacent R wave peaks) are obtained according to the adjacent R wave peak positions, and a heart rate variability feature acquired by the electrocardiographic signal is obtained according to the RR intervals.

More specifically, the following time-domain heart rate variability indexes can be acquired as the time-domain indexes of the heart rate variability features: SDNN (standard deviation of RR intervals of all sinus heartbeats), RMSSD (root-mean-square of difference of adjacent RR intervals, root-mean-square of difference of RR intervals), and AVNN (average of all RR intervals), which are heart rate analysis in the time domain, wherein each index contains information of regulation of neurohumoral variability factors on cardiovascular system, thereby being able to judge the condition and prevention of cardiovascular diseases and the like.

In this embodiment, the heart beat is a process from the beginning of one heart beat to the beginning of the next heart beat, and the cutting is performed by taking the heart beat as a unit, and each heart beat in the electrocardiographic signal is divided into a heart beat segment, that is, the heart beat segment is the electrocardiographic signal corresponding to the heart beat. The process of segmenting the heartbeat segment is specifically as follows, peak positions of all R waves in the electrocardiographic signal can be obtained, and then RR intervals adjacent to each R wave, that is, intervals between peaks, specifically, for a certain R wave, a first RR interval of the R wave and a previous adjacent R wave and a second RR interval of the R wave and a next adjacent R wave are obtained, as shown in fig. 2, a rear 1/3 part of the first RR interval and a front 2/3 part of the second RR interval are cut out as heartbeat segments corresponding to the R wave. Of course, the division of the heart beat segments is not limited to the above-mentioned method, and other methods are also possible. It should be noted that the heart rate variability features and the heart beat segments are not acquired sequentially.

S103, acquiring heart beat characteristics of each heart beat segment to form a heart beat characteristic group.

In this embodiment, the cardiac characteristics may be extracted in the following manner, first, high and low frequency noises in the electrocardiographic signal of the cardiac segment are filtered out, and considering that the electrocardiographic acquisition device may be subjected to various interferences in the actual use process, which reduces the recognizable degree of the electrocardiographic signal, where the interferences include: high-frequency electromyographic interference, power frequency interference, low-frequency baseline drift and the like. Specifically, high and low frequency noises in the electrocardiosignals can be respectively filtered out through wavelet multi-resolution analysis and a Butterworth filter;

then, carrying out normalization processing on the filtered electrocardiosignals, wherein the electrocardiosignal values are mapped between 0 and 1, the traditional data normalization processing can be adopted, the process is as follows, all signal maximum values and signal minimum values of the electrocardiosignals of the segment are taken, and then, the following operation is carried out on each sampling point, (the sampling point signal instantaneous value-signal minimum value)/(the signal maximum value-signal minimum value), so that all the sampling points of the segment are mapped between 0 and 1; the normalization has the significance that for a machine learning algorithm, the normalization enables data to keep good distribution, a model algorithm can be more easily converged, and the application range of the model is enlarged. And finally, inputting the electrocardiosignals of the heart beat fragments subjected to normalization processing into a heart beat feature extraction model trained in advance, and extracting heart beat features.

Furthermore, the heartbeat features of the heartbeat segments are combined to form a heartbeat feature group, and because the electrocardiographic signal includes a plurality of heartbeat segments, a plurality of features with different dimensions can be obtained from each heartbeat segment, and for the heartbeat feature group obtained by combining a large number of heartbeat features, the heartbeat feature group is subjected to dimensionality reduction by using a Principal Component Analysis (PCA) method to compress data volume.

And S104, inputting the heart rate variability features and the heart beat feature group into an atrial fibrillation recognition model to obtain an atrial fibrillation recognition result.

Since the most significant features of atrial fibrillation are obvious irregularities of RR intervals, index of heart rate variability reflects the change condition of successive heart cycle differences, the two are actually related, heart beat is the basic component of heart rhythm, and it is the most essential practice to extract heart beat features, the heart rate variability features and the heart beat features are used as evaluation parameters of atrial fibrillation in the embodiment. In this embodiment, the heart beat features after the heart rate variability features are dimensionality reduced are input into a pre-trained atrial fibrillation recognition model, so that an atrial fibrillation recognition result is obtained, that is, the recognition of atrial fibrillation in the electrocardiosignal is completed.

According to the atrial fibrillation identification method provided by the embodiment, the peak position of the R wave in the electrocardiosignal to be identified is obtained, then the heart beat characteristics of each heart beat section are obtained to form a heart beat characteristic group according to the heart beat variability characteristics and the heart beat sections of the electrocardiosignal to be identified, which are determined according to the peak position of the R wave, and finally the heart beat variability characteristics and the heart beat characteristics are input into an atrial fibrillation identification model to obtain an atrial fibrillation identification result. Because the atrial fibrillation recognition model is obtained through mass data training and is a model among heart rate variability characteristics, heart beat characteristics and heart rhythm types, the atrial fibrillation recognition model can effectively recognize atrial fibrillation by inputting the heart rate variability characteristics and the heart beat characteristics of electrocardiosignals into the pre-trained atrial fibrillation recognition model, and accordingly reliability and accuracy of atrial fibrillation recognition are improved.

Based on the above embodiment, as shown in fig. 3, the acquiring the heart beat feature of each heart beat segment in S103 may specifically include:

s201, filtering high-frequency and low-frequency noises in the electrocardiosignals of the heartbeat segment.

In this embodiment, high and low frequency noise in the electrocardiographic signal is filtered out of the heartbeat segment, which also avoids reduction of signal identifiability caused by various interferences in the use process of the electrocardiographic acquisition device. Specifically, high and low frequency noises in the electrocardiosignals can be respectively filtered through wavelet multi-resolution analysis and a Butterworth filter. More specifically, a Db4 wavelet basis function and 8-order wavelet transform can be selected to perform multi-scale decomposition on the intercepted heart beat, and the high-frequency part in the heart beat signal is filtered by performing soft threshold processing and reconstruction on the wavelet coefficient; then, a butterworth filter is selected to filter out the low frequency part of the heartbeat signal. Where Db4 refers to one of many wavelet basis functions, and the 8-order wavelet transform refers to the number of times that the wavelet-transformed signal is subjected to wavelet transformation. Specifically, wavelet decomposition is carried out on the electrocardiosignals containing noise, the number of layers of decomposition is 8, soft threshold processing is carried out on the high-frequency coefficients of the wavelet decomposition (when the absolute value of the wavelet coefficients is more than or equal to a given threshold, the value of the wavelet coefficients is reduced by the threshold, and when the absolute value of the wavelet coefficients is less than the threshold, the value of the wavelet coefficients is 0), and then one-dimensional reconstruction is carried out according to the wavelet coefficients of the first order to the eighth order obtained by the wavelet decomposition to obtain the electrocardiosignals with the high-frequency parts removed; and finally, designing a Butterworth filter to filter out the low-frequency part in the electrocardiosignal.

S202, normalization processing is carried out on the filtered electrocardiosignals.

In this embodiment, the filtered ecg signal is normalized and the value of the ecg signal is mapped to a value between 0 and 1. The traditional data normalization processing can be adopted, the process is as follows, the maximum value and the minimum value of all signals of the electrocardiosignal of the segment are taken, and then the following operation is carried out on each sampling point, (the instantaneous value of the signal of the sampling point-the minimum value of the signal)/(the maximum value of the signal-the minimum value of the signal), so that all the sampling points of the segment are mapped between 0 and 1; the normalization has the significance that for a machine learning algorithm, the normalization enables data to keep good distribution, a model algorithm can be more easily converged, and the application range of the model is enlarged.

S203, inputting the electrocardiosignals subjected to normalization processing into a heart beat feature extraction model to obtain heart beat features.

In this embodiment, the electrocardiographic signal subjected to normalization processing is input to a heart beat feature extraction model trained in advance, and heart beat features are extracted.

In the embodiment, the high-frequency and low-frequency noises in the electrocardiosignals are filtered out through the heartbeat segment, so that the reduction of the signal identification degree caused by various interferences in the use process of the electrocardio acquisition equipment is avoided; normalization processing is carried out on the filtered electrocardiosignals, so that data keep good distribution, a model algorithm can be more easily converged, and the application range of the model is enlarged; the electrocardiosignals after normalization processing are input into a heart beat feature extraction model to obtain heart beat features, so that the obtained heart beat features are more accurate.

In the process of establishing the heart beat feature extraction model shown in fig. 4, the specific steps may be as follows:

s301, first electrocardiosignal samples of different heart beat types are obtained.

Wherein the heart beat types include a normal heart beat, a left bundle branch block, a right bundle branch block, an atrial premature beat, a ventricular premature beat, and a paced heart beat.

In this embodiment, the heartbeat data (cardiac signal) with identification information may be obtained from the MIT arrhythmia data set as a training sample, i.e., a first cardiac signal sample. Wherein the identification information comprises a normal heart beat (N), a left bundle branch block (L), a right bundle branch block (R), an atrial premature beat (A), a ventricular premature beat (P), and a pacing heart beat (V).

Further, in order to ensure the balance of the distribution of each type of data and to make the network easy to converge and optimize, a random elimination method can be adopted to adjust the proportion of the first electrocardiosignal samples of each heartbeat type, specifically, a large number of normal heartbeats can be eliminated at random, and the strategy is as follows: setting threshold probability, generating a random number within the range of 0-1 when the heart beat type is detected to be a large number of normal heart beats, if the random number is greater than the threshold, retaining the heart beat segment, otherwise, discarding the random number, thereby realizing the proportion balance of various types of sample data. For example, in the original dataset, 200 normal heart beats, 100 atrial premature beats; the normal heart beat is large in number, a threshold value is set to be 0.5, when the normal heart beat is extracted, a random number is generated, if the random number is larger than 0.5, the normal heart beat is reserved, otherwise, the random number is discarded, and the final normal heart beat number is basically consistent with the small number of the atrial premature heart beats.

S302, preprocessing each first electrocardiosignal sample.

Wherein the pre-processing comprises: and dividing the first electrocardiosignal sample into heart beat fragments, and filtering and normalizing high and low frequency noise.

In this embodiment, for each first cardiac signal sample, the cutting is performed first by taking the heart beat as the unit, and the cutting method is the same as S102. After completion of the cutting, the largest RR interval, as shown in fig. 5a, was selected as the length of the heart beat sample for training, and 0 was added to the head and tail of the heart beat segment obtained by cutting so that the lengths of the heart beat segment samples were the same, as shown in fig. 5 b.

S303, establishing a network structure of the convolutional neural network, inputting each preprocessed first electrocardiosignal sample into the convolutional neural network for training, and generating a heart beat classification model.

In this embodiment, a network structure of a convolutional neural network as shown in fig. 6 is established to realize the identification of the heartbeat type, and a network model and results thereof are as follows, wherein a first layer is an input layer, a seventh classification layer is an output layer, and other intermediate layers are hidden layers. In addition, the invention saves the learning parameters of the model so as to extract the heart beat characteristics for identifying atrial fibrillation by utilizing the last hidden layer (namely, the full connection layer) of the model.

(1) Input layer (Input): 1024 pieces of electrocardiogram data are input each time in a batch processing mode for one-dimensional data after heart beat data preprocessing.

(2) Convolutional layer (Conv): one-dimensional convolution is adopted, the length of a convolution kernel is 5, the number of input channels is 1, the number of output channels is 32, and the input and output sizes are unchanged by filling 0 from beginning to end.

(3) BN layer (Batch Normalization): and (5) marking layers in batches, and recording the mean value and variance of each layer in the training process, wherein the attenuation coefficient is 0.997.

(4) Active layer (ReLU): a ReLU nonlinear activation function is employed.

(5) Residual Block (Residual Block): the length of a convolution kernel in the convolution layer is 5, and the input and output channels are 32; the residual part adopts a summation mode; the Pooling layer uses the Max Pooling method (Max Pooling), with a scale of 5 and a sliding distance of 2. Here, the residual block is superimposed 5 times.

(6) Fully connected layer (Dense): the number of the neurons output by the full connection layer is 32.

(7) Classification layer (Softmax): the conclusion is output for each sample, and the total is classified into 6 types.

Optimizing the model parameters by using an Adam optimization method, wherein the step length is 0.001; the moment estimated attenuation rates are respectively rho₁＝0.9，ρ₂0.999. And setting the batch size to be 1024, the loss function to be cross entropy and the iteration period to be 5000 times.

By training and optimizing the model parameters, a confusion precision matrix as in fig. 7 is obtained. And the average accuracy of the model parameters on the test set is verified to reach 97.9%.

And for the features extracted from the last hidden layer, performing visualization processing on the features by using t-SNE, as shown in FIG. 8.

As can be seen by the t-SNE clustering algorithm, different heartbeats are well distinguished and represented in the last hidden layer of the model.

S304, acquiring a heart beat feature extraction model according to the heart beat classification model.

In the present embodiment, the above-described model except for the output layer is saved as a heart beat feature extraction model.

Specifically, because the heartbeat classification model is obtained by training the first cardiac signal sample in the MIT arrhythmia data set, potential representations of the interior of heartbeats are gradually obtained, and heartbeat features are continuously obtained in continuous learning, it is possible that different types of heartbeats of the last full connection layer (i.e., the last hidden layer) have different data distributions (data characteristics) on the layer, which means that the heartbeat classification model is not only suitable for classification in the MIT arrhythmia data set, but also has certain description capability on heartbeats themselves. If the heart beat classification model is used for direct output, the model can be used for predicting the MIT heart beat, namely classifying the heart beat; however, if the classification layer is not applied, only the hidden layer is applied, the heart beat features can be obtained, and therefore the heart beat feature extraction model is obtained.

On the basis of the above embodiment, as shown in fig. 9, the embodiment further provides a process for creating an atrial fibrillation recognition model, and the specific steps may be as follows:

s401, second electrocardiosignal samples of different heart rhythm types are obtained.

Wherein the types of heart rhythms include normal heart rhythms, atrial fibrillation heart rhythms, and other heart rhythms, wherein the other heart rhythms are arrhythmia except for atrial fibrillation heart rhythms.

In this embodiment, labeled cardiac rhythm data in the MIT arrhythmia data set is used to train an atrial fibrillation recognition model. Specifically, electrocardiosignal samples of three types of normal heart rhythm, atrial fibrillation heart rhythm and other heart rhythms are extracted from the MIT arrhythmia data set and serve as the second electrocardiosignal sample, wherein the other heart rhythms are other arrhythmia except for atrial fibrillation heart rhythm, such as atrial bigeminy rhythm, ventricular autonomic rhythm and other heart rhythms, so that the practical usability of the algorithm is guaranteed.

Further, for MIT arrhythmia data set data, compared with other two types, atrial fibrillation heart rate electrocardiosignal samples are less, in order to guarantee model training quality and optimization effect, the heart beat is used as a basic unit, data enhancement is performed on atrial fibrillation signals in a random sliding mode, namely, a plurality of atrial fibrillation heart rate fragments are intercepted from second electrical signal samples of atrial fibrillation heart rates in a random sliding mode to serve as second electrical signal samples of new atrial fibrillation heart rates, and the strategy is as follows:

assuming that each atrial fibrillation rhythm segment contains k heart beats, the starting point at which the atrial fibrillation waveform occurs is B, and the ending point at the end of the atrial fibrillation is E, then the original rhythm segment (E-B)/k heart beats are obtained, as shown in FIG. 10. Randomly taking a plurality of integers between B, E, and if the integers are equal to the starting point of the existing segment, discarding the segment; otherwise, taking the heart beat as a starting point, cutting k heart beats after the starting point as a new atrial fibrillation sample, and discarding the heart beats between the starting point and E if the number of the heart beats is less than k, thereby enhancing the atrial fibrillation data set.

S402, acquiring heart rate variability characteristics and a heart beat characteristic group of each second electrocardiosignal sample.

In this embodiment, for each second electrocardiographic signal sample, the heart rate variability feature may be obtained by the same method as that in S102, that is, by obtaining peak positions of all R waves in the electrocardiographic signal, a series of RR intervals are obtained according to adjacent R peak positions, and the heart rate variability feature obtained by the electrocardiographic signal is obtained according to the RR intervals. More specifically, the following time-domain heart rate variability indexes can be acquired as the time-domain indexes of the heart rate variability features: standard deviation of all RR intervals of SDNN; root mean square of RMSSD adjacent RR interval differences; average of all RR intervals of AVNN.

In the present embodiment, the heart rhythm is composed of a plurality of heart beats, and based on this, in addition to extracting the statistical characteristics of the heart rhythm segment, the present invention will also migrate to obtain the characteristics of each heart beat in the heart rhythm segment of each second cardiac electrical signal sample by using the previously trained heart beat characteristic extraction model.

Specifically, heart beat segments are segmented for each second electrocardiosignal sample, and high-low frequency noise filtering and normalization processing are carried out; and inputting the heart beat characteristic into the heart beat characteristic extraction model, and extracting heart beat characteristics to obtain the heart beat characteristics of the second electrocardiosignal sample.

The heart beat segmentation, the high-low frequency noise filtering and the normalization are the same as those of the embodiment, the heart beat set of each second electrocardiosignal sample is obtained, the heart beat feature extraction model is combined to obtain the intrinsic heart beat feature of each heart beat, and the new representation of each heart beat is obtained through the last 32 neuron full-junction layers. And rearranging the heart beat characteristics into a row, and if each second cardiac signal sample heart rate segment contains 20 heart beats, taking the total number of the heart beat characteristics of the heart rate segment as 32 multiplied by 20 as the automatically extracted heart beat characteristic group in the heart rate segment.

For such a large number of feature numbers, the Principal Component Analysis (PCA) may be used to perform dimensionality reduction processing on the feature numbers in this embodiment. Specifically, heart rhythm segments in all second cardiac signal samples are cut into respective heart rhythm sets, the features of each heart rhythm are obtained through the last layer of a heart rhythm model, the heart rhythm features from one heart rhythm set are rearranged into a row, heart rhythm feature sets corresponding to a plurality of heart rhythm segments are merged, and on the basis of the row features, mean value processing, covariance feature values and feature vector solving are carried out to obtain a new low-dimensional representation of the specified feature retention degree. And recording the feature mean value and the conversion matrix in the PCA conversion process so as to be used when the dimension of the test data is reduced, and combining the heart rate variability indexes extracted from the heart rate segments before to be used as the features of the heart rate segments for the next training process, wherein the heart rate feature extraction process is shown in FIG. 11. It should be noted that, if the heart beat feature set is subjected to dimension reduction processing during the training of the atrial fibrillation recognition model, correspondingly, the heart beat feature set may also be subjected to dimension reduction processing before the heart beat variability feature and the heart beat feature set are input into the atrial fibrillation recognition model when the atrial fibrillation recognition model is applied in S104, and the dimension reduction processing result is input into the atrial fibrillation recognition model as the heart beat feature set.

S404, training the heart rate variability features and the heart beat feature group of the second cardiac signal sample to generate the atrial fibrillation recognition model.

In this embodiment, the heart rate variability characteristics of the normal heart rhythm, atrial fibrillation heart rhythm, and other heart rhythms in the second set of cardiac signal samples, and the heart beat characteristics are used as training data. And establishing a 2-layer full-connection network structure which comprises a full-connection layer and a classification layer.

The number of neurons of the full connection layer is ensured to be the same as that of neurons of the input layer (dimensionality after PCA dimension reduction + HRV characteristic dimensionality), the number of neurons of the output layer is 3, and Softmax is adopted as a classification function.

The training data and the test data of the present embodiment are both from the MIT data set, and k is 5,10,20 for the number of heart beats k of each heart rhythm segment; the retention contribution rates to the PCA are respectively 90%, 99% and 99.9%; the parameters of the Adam method are the same as those of the learning strategy sampling method.

The indexes of the atrial fibrillation model for evaluating the performance of the model adopt the indexes of sensitivity, specificity and accuracy, and the calculation formulas of the three indexes are as follows:

wherein K represents the type of the heart rhythm, and TP represents the number of correctly classified types H; TN represents the number of non-H classes not classified into H classes; FP represents the number of non-H-class errors classified into H-class; FN indicates the number of classes into which the H class is classified. Its specific meaning is exemplified by normal heart rhythm, as shown in table 1.

TABLE 1

After repeated tests, it can be concluded that table 2 shows:

TABLE 2

From this, it is known that 20 beats are selected for each heart rate, 90% of PCA remains, and the best recognition accuracy is obtained, SPE 98.2%, SEN 97.3%, and ACC 98.0%.

Further, this embodiment also provides an overall framework as shown in fig. 12, which includes heart beat data preprocessing, a heart beat classification model, heart beat data preprocessing, feature extraction and a heart rhythm classification model, for the electrocardiographic raw data is subjected to heart rhythm data preprocessing to obtain a heart rhythm data set, heart beat data preprocessing is called, then feature extraction is performed, including extraction of heart beat variability features and extraction of heart beat features, where heart beat feature extraction calls the heart beat classification model (since the heart beat feature extraction model is a part of the heart beat classification model, this is equivalent to obtaining a heart beat feature group by the heart beat feature extraction model), the heart beat feature group and the heart beat variability features are input into the atrial fibrillation recognition model together for heart rhythm classification after feature dimensionality reduction, so as to determine whether the heart rhythm type is atrial fibrillation, and the specific processing procedure of the framework has been described in detail in the above embodiment, and will not be described in detail herein.

The atrial fibrillation identification method provided by the embodiment is combined with a deep learning method, and provides a new strategy for detecting atrial fibrillation in real time and with high precision and finding the atrial fibrillation in time. Specifically, the heart beat features of the heart rhythm to be recognized are extracted through transfer learning, and the high-precision detection of atrial fibrillation is realized by combining the traditional heart rate variability features. Compared with the traditional signal identification method, the method has higher identification precision and reliability. Meanwhile, as a plurality of patients with atrial fibrillation are paroxysmal, the atrial fibrillation identification method in the embodiment can be applied to a monitoring device, so that atrial fibrillation with interval attacks can be conveniently monitored, and the workload of electrocardio monitoring is reduced.

Fig. 13 is a block diagram of an atrial fibrillation recognition device according to an embodiment of the present invention. The atrial fibrillation recognition apparatus provided by this embodiment may execute the processing flow provided by the atrial fibrillation recognition method embodiment, as shown in fig. 13, the atrial fibrillation recognition apparatus includes a peak position acquisition module 61, a heart rate variability feature acquisition module 62, a heartbeat segment acquisition module 63, a heartbeat feature acquisition module 64, and an atrial fibrillation recognition module 65.

The peak position acquiring module 61 is used for acquiring the peak position of the R wave in the electrocardiosignal to be identified;

a heart rate variability feature acquisition module 62, configured to determine a heart rate variability feature according to a peak position of the R wave;

a heartbeat segment obtaining module 63, configured to determine heartbeat segments according to peak positions of the R waves;

a heart beat feature acquisition module 64, configured to acquire heart beat features of each heart beat segment to form a heart beat feature group;

and the atrial fibrillation recognition module 65 is configured to input the heart rate variability features and the heart beat feature group into an atrial fibrillation recognition model to obtain an atrial fibrillation recognition result.

Further, the heart rate variability feature acquisition module 62 is configured to:

obtaining RR intervals according to the peak positions of adjacent R waves;

acquiring a heart rate variability feature from all of the RR intervals.

Further, the heartbeat segment acquiring module 63 is configured to:

for the peak position of any R wave in the electrocardiosignals to be identified, acquiring a first RR interval between the peak position of any R wave in the electrocardiosignals to be identified and the peak position of the previous adjacent R wave and a second RR interval between any R wave in the electrocardiosignals to be identified and the peak position of the next adjacent R wave;

Further, the heartbeat feature acquisition module 64 is configured to:

normalizing the filtered electrocardiosignals;

Further, the heartbeat feature acquisition module 64 is further configured to:

after first electrocardiosignal samples of different heart beat types are obtained, the proportion of the first electrocardiosignal samples of each heart beat type is adjusted by adopting a random elimination method for the first electrocardiosignal samples.

Further, the atrial fibrillation identification module 65 is further configured to:

segmenting heartbeat segments for each second electrocardiosignal sample, and carrying out high-low frequency noise filtering and normalization processing; and inputting the heart beat characteristic into the heart beat characteristic extraction model, and extracting heart beat characteristics to obtain a heart beat characteristic group of the second electrocardiosignal sample.

The atrial fibrillation recognition apparatus provided in this embodiment may be specifically configured to perform the method embodiments provided in fig. 1, 3, 4 and 9, and specific functions will not be described herein again.

According to the atrial fibrillation recognition device provided by the embodiment, the peak position of an R wave in an electrocardiosignal to be recognized is obtained, and then the heart rate variability feature and a heart beat segment are determined according to the peak position of the R wave; acquiring heart beat characteristics of each heart beat segment to form a heart beat characteristic group; and finally, inputting the heart rate variability characteristics and the heart beat characteristic group into an atrial fibrillation recognition model to obtain an atrial fibrillation recognition result. According to the atrial fibrillation recognition method, the heart rate variability characteristics and the heart beat characteristics of the electrocardiosignals are input into the pre-trained atrial fibrillation recognition model, so that atrial fibrillation can be effectively recognized, and the reliability and the accuracy of atrial fibrillation recognition are improved.

Fig. 14 is a block diagram of an atrial fibrillation identification device according to another embodiment of the present invention. As shown in fig. 14, the present embodiment provides an atrial fibrillation identifying device including: a processor 71; a memory 72; and a computer program.

Wherein the computer program is stored in the memory 72 and configured to be executed by the processor 71 to implement the processing procedure provided by the method embodiments provided in fig. 1, 3, 4 and 9, and the specific functions are not described herein again.

More specifically, the atrial fibrillation identifying device further includes a receiver 73 and a transmitter 74, wherein the receiver 73, the transmitter 74, the processor 71, and the memory 72 are connected by a bus.

Another embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon;

which when executed by a processor implements the atrial fibrillation recognition method according to the above-described embodiments.

In the embodiments provided in the present invention, it should be understood that the disclosed apparatus and method may be implemented in other ways. For example, the above-described apparatus embodiments are merely illustrative, and for example, the division of the units is only one logical division, and other divisions may be realized in practice, for example, a plurality of units or components may be combined or integrated into another system, or some features may be omitted, or not executed. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.

The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

In addition, functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may exist alone physically, or two or more units are integrated into one unit. The integrated unit can be realized in a form of hardware, or in a form of hardware plus a software functional unit.

The integrated unit implemented in the form of a software functional unit may be stored in a computer readable storage medium. The software functional unit is stored in a storage medium and includes several instructions to enable a computer device (which may be a personal computer, a server, or a network device) or a processor (processor) to execute some steps of the methods according to the embodiments of the present invention. And the aforementioned storage medium includes: various media capable of storing program codes, such as a usb disk, a removable hard disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a magnetic disk, or an optical disk.

It is obvious to those skilled in the art that, for convenience and simplicity of description, the foregoing division of the functional modules is merely used as an example, and in practical applications, the above function distribution may be performed by different functional modules according to needs, that is, the internal structure of the device is divided into different functional modules to perform all or part of the above described functions. For the specific working process of the device described above, reference may be made to the corresponding process in the foregoing method embodiment, which is not described herein again.

Finally, it should be noted that: the above embodiments are only used to illustrate the technical solution of the present invention, and not to limit the same; while the invention has been described in detail and with reference to the foregoing embodiments, it will be understood by those skilled in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some or all of the technical features may be equivalently replaced; and the modifications or the substitutions do not make the essence of the corresponding technical solutions depart from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An atrial fibrillation identification device, comprising: the device comprises a peak position acquisition module, a heart rate variability characteristic acquisition module, a heart beat fragment acquisition module, a heart beat characteristic acquisition module and an atrial fibrillation identification module;

the peak position acquiring module is used for acquiring the peak position of the R wave in the electrocardiosignal to be identified;

the heartbeat feature acquisition module is used for acquiring heartbeat features of each heartbeat segment to form a heartbeat feature group;

the atrial fibrillation recognition module is used for inputting the heart rate variability features and the heart beat feature group into an atrial fibrillation recognition model to obtain an atrial fibrillation recognition result;

the atrial fibrillation identifying module is specifically used for acquiring second cardiac electrical signal samples of different heart rhythm types, wherein the heart rhythm types comprise a normal heart rhythm, an atrial fibrillation heart rhythm and other heart rhythms, and the other heart rhythms are other arrhythmia except for the atrial fibrillation heart rhythm; acquiring heart rate variability characteristics and a heart beat characteristic group of each second electrocardiosignal sample; and training the heart rate variability features and the heart beat feature group of the second cardiac signal samples to generate the atrial fibrillation recognition model.

2. The device according to claim 1, characterized in that said heart rate variability feature acquisition module, in particular for acquiring RR intervals from the peak positions of adjacent R-waves; acquiring a heart rate variability feature from all of the RR intervals.

3. The apparatus according to claim 1, wherein the heartbeat segment acquiring module is specifically configured to acquire, for a peak position of any one R wave in the electrocardiographic signal to be identified, a first RR interval between the peak position of the R wave and a previous adjacent R wave, and a second RR interval between the peak position of the R wave and a next adjacent R wave; and cutting the rear 1/3 part of the first RR interval and the front 2/3 part of the second RR interval to be used as the heart beat segment corresponding to the R wave.

4. The device according to claim 1 or 3, wherein the heartbeat feature acquisition module is specifically configured to filter out high and low frequency noise in the electrocardiographic signal for the heartbeat segment; normalizing the filtered electrocardiosignals; and inputting the electrocardiosignals subjected to normalization processing into a heart beat feature extraction model to obtain heart beat features.

5. The apparatus according to claim 4, wherein the heartbeat feature acquisition module is specifically configured to acquire first cardiac signal samples of different heartbeat types, the heartbeat types including a normal heartbeat, a left bundle branch block, a right bundle branch block, an atrial premature beat, a ventricular premature beat, and a paced heartbeat; preprocessing each of the first cardiac signal samples, wherein the preprocessing comprises: dividing the first electrocardiosignal sample into heart beat fragments, and filtering and normalizing high and low frequency noise; establishing a network structure of a convolutional neural network, inputting each preprocessed first electrocardiosignal sample into the convolutional neural network for training, and generating a heartbeat classification model; and acquiring a heart beat feature extraction model according to the heart beat classification model.

6. The apparatus of claim 1, wherein the atrial fibrillation identification module is further configured to: and intercepting a plurality of atrial fibrillation rhythm segments from the second cardiac signal samples of the atrial fibrillation rhythm in a random sliding mode to serve as second cardiac signal samples of the new atrial fibrillation rhythm.

7. The apparatus of claim 1, wherein the atrial fibrillation identification module is further configured to: segmenting heartbeat segments for each second electrocardiosignal sample, and carrying out high-low frequency noise filtering and normalization processing; and inputting the heart beat characteristic into a heart beat characteristic extraction model, and extracting heart beat characteristics to obtain a heart beat characteristic group of the second electrocardiosignal sample.

8. The apparatus of claim 7, wherein the atrial fibrillation recognition module is further configured to perform dimension reduction on the heart beat feature set before inputting the heart rate variability features and the heart beat feature set into an atrial fibrillation recognition model, and taking a dimension reduction result as the heart beat feature set; before the heart rate variability features and the heart beat feature group of the second electrocardiosignal sample are trained, performing dimensionality reduction processing on the heart beat feature group of the second electrocardiosignal sample, and taking a dimensionality reduction processing result as the heart beat feature group of the second electrocardiosignal sample.