CN109117730B

CN109117730B - Real-time electrocardiogram atrial fibrillation judgment method, device and system and storage medium

Info

Publication number: CN109117730B
Application number: CN201810760168.0A
Authority: CN
Inventors: 刘成良; 史浩天; 张飞; 刘金磊
Original assignee: Shanghai Xiaxian Electromechanical Science And Technology Development Co ltd
Current assignee: Shenzhen Hart Intelligent Technology Co.,Ltd.
Priority date: 2018-07-11
Filing date: 2018-07-11
Publication date: 2021-05-11
Anticipated expiration: 2038-07-11
Also published as: CN109117730A

Abstract

An electrocardiogram atrial fibrillation real-time judgment method, device and system and a computer storage medium thereof can analyze an electrocardiogram in real time based on wearable equipment. The method extracts features such as RR interphase, root mean square, wavelet coefficient and the like, fully utilizes hidden information in the waveform, and combines rule judgment and a machine learning model to judge atrial fibrillation. And finally, fusing the judgment results of the plurality of models, and giving a score of the atrial fibrillation, namely the doubtful degree of the atrial fibrillation. The invention only needs to collect 20s electrocardiogram and can accurately analyze and judge the atrial fibrillation in real time.

Description

Real-time electrocardiogram atrial fibrillation judgment method, device and system and storage medium

Technical Field

The invention relates to the field of electrocardiogram monitoring, in particular to an electrocardiogram atrial fibrillation real-time judgment method, device and system and a computer storage medium thereof.

Background

In recent years, cardiovascular diseases have become more prevalent worldwide, and the number of deaths due to cardiovascular diseases has also increased. The electrocardiogram is an effective method for observing the potential change of the heart and judging cardiovascular diseases. However, since the change in the amplitude and frequency components of the electrocardiogram waveform is slight, it is difficult and time-consuming for a doctor to determine cardiovascular diseases by electrocardiogram. In addition, long-term visual observation may also lead to misdiagnosis. Therefore, computer-aided electrocardiogram (ecg) determination is becoming more and more important, and has been studied and applied in a large amount. Which can significantly improve the diagnosis of heart disease at an acceptable price.

Atrial fibrillation is a common and important supraventricular arrhythmia, especially common in the elderly population. Atrial fibrillation is a disorder of atrial rhythm resulting from many small reentrant loops caused by atrial-dominated reentrant loops, which occur with disorganized activation and ineffective contraction of the atria. Which is a major hazard to the patient. Patients often show palpitation and weakness due to extremely uneven ventricular beats. Atrial fibrillation is also found in all patients with organic heart disease, has a high rate of onset and a long duration, and may deteriorate cardiac function, causing serious complications such as heart failure and arterial embolism, resulting in disability or increased mortality of the patient. Therefore, an intelligent judgment method capable of timely and accurately judging atrial fibrillation is needed.

At present, in addition to the conventional method for detecting the electrocardiogram at the hospital and judging by the doctor, for paroxysmal atrial fibrillation, a 24-hour hold is often adopted to continuously collect the electrocardiogram, and the data of 2-3 days are transmitted to the hospital and judged by the doctor. The existing computer-aided judgment method usually needs to collect signals for a certain time and then analyze and judge the signals through a computer. Portable hardware-based approaches, which have emerged, also require hundreds of waveforms to be collected for a few minutes to make a determination. But many also require implantable based detection devices. In the existing atrial fibrillation determining method, the determination of atrial fibrillation is usually based on information of RR intervals, such as the type, size change, turning point, etc. of RR intervals. And analyzing by adopting one or more judgment conditions, and finally giving a judgment on whether the atrial fibrillation exists.

1. The existing atrial fibrillation judgment scheme usually only uses the absolute irregularity of RR intervals as a basis to make a relevant criterion. This results in the waveform not being used to include other information useful in determining atrial fibrillation. The accuracy of the final judgment result is not sufficiently influenced by the insufficient utilization of the hidden disease information.

2. The existing method, no matter collecting a certain amount of data and seeing the data for doctors, or using computer analysis, needs a long time to obtain the analysis result. Existing portable hardware-based methods also require long acquisition times. Therefore, the method cannot be rapidly judged in real time.

3. The existing atrial fibrillation judging method can judge whether atrial fibrillation exists or not when the judgment is finished. Giving an absolute judgment under the condition that absolute accuracy cannot be guaranteed can affect the judgment of the patient on the disease. Existing methods do not give the possibility of having atrial fibrillation.

Disclosure of Invention

The invention aims to solve the technical problems that electrocardiogram waveform information cannot be fully utilized, time delay or time consumption is long in judgment, and judgment results are too great, and the like, and provides an electrocardiogram atrial fibrillation real-time judgment method, device and system and a computer storage medium thereof, which can quickly realize real-time atrial fibrillation monitoring and judgment through 20s of electrocardiogram.

The invention is realized by the following technical scheme, and a first aspect of the invention provides a real-time judgment method for atrial fibrillation by electrocardiogram, which comprises the following steps:

step 1, electrocardiogram preprocessing, namely obtaining an available electrocardiogram marked with waveform characteristics and electrocardiogram wave bands for segmenting the available electrocardiogram;

step 2, inputting the available electrocardiogram marked with the waveform characteristics into a rule model, and outputting a rule model result after the rule model is judged based on a ventricular escape medical judgment rule; inputting the segmented electrocardiogram wave bands into a machine learning model trained through hospital data, and obtaining a machine learning model result after the machine learning model is judged;

and 3, performing fusion calculation on the rule model result and the machine learning model result to obtain a final result of the doubtful degree of atrial fibrillation.

In a specific embodiment, the electrocardiogram preprocessing of step 1 includes the following steps:

step 11, filtering the electrocardiogram collected originally to remove signal noise and obtain an available electrocardiogram;

step 12, carrying out waveform identification on the available electrocardiogram, and identifying and marking P waves and QRS wave groups in the electrocardiogram signals;

and step 13, carrying out waveform segmentation on the labeled available electrocardiogram, and dividing the waveform into certain wave bands which can be used for machine learning to obtain electrocardiogram wave bands.

In a specific embodiment, the filtering is performed by using a wavelet threshold method to perform denoising, a db4 wavelet is used to decompose 8 layers of signals, the decomposed wavelet coefficients are processed by a soft threshold method to obtain new wavelet coefficients, and signal reconstruction is performed by using the new wavelet coefficients to obtain filtered electrocardiosignals which are the available electrocardiograms;

and/or the waveform identification is based on B-spline biorthogonal wavelet detection QRS wave complex, and the position of Q, R, S points is determined; detecting a P wave based on first-order difference, and determining the position of a P point;

and/or the waveform segmentation takes the position 0.3s before the R wave crest as a starting point and the position 0.3s after the R wave crest as an end point, so that the starting point is taken as a heart beat, and 6 heart beats are divided into a wave band to be taken as an electrocardiogram wave band sample to be input into the machine learning model.

In a specific embodiment, the rule model judgment in step 2 includes the following steps:

step A21, calculating RR interphase, P-wave height and R-wave height according to the input available electrocardiogram;

step A22, detecting abnormal values in RR intervals and removing the abnormal values;

step A23, obtaining a first model result S1, a second model result S2, a third model result S3 and a fourth model result S4 respectively through a first rule model for judging the maximum value and minimum value ratio of RR intervals, a second rule model for judging the deviation of RR intervals and a mean value, a third rule model for judging the number of RR interval groups which accord with complete compensation gaps and approximate premature beat types, and a fourth rule model for judging the waveform ratio of the P-wave R-wave height ratio in a threshold range;

step A24, determine S1 equal to 0: if yes, S is 0; if not, S is S1+ S2-S3-S4;

and step A25, outputting S.

In a specific embodiment, the abnormal value detection in step a22 is used to determine whether there is an abnormal value in the RR intervals, and first, a mean value of all RR intervals is calculated; and judging whether the value of each RR interval is more than 0.5 time of the mean value and less than 1.6 times of the mean value, if not, exceeding the threshold range, determining that the RR interval is an abnormal value and removing the abnormal value.

In a specific embodiment, the rule model one in step a23 includes:

(1) inputting the ratio q of the maximum value to the minimum value of all RR intervals in the available electrocardiogram, and initializing a parameter d to make d equal to 0;

(2) and (3) judging the q value:

if q is less than or equal to 1.1, the parameter d is 5;

if q is more than 1.1 and less than or equal to 1.9, calculating the parameter d as-4.745 xq + 10.2195;

if q is more than 1.9 and less than or equal to 2.1, the parameter d is 1.204;

if q is more than 2.1 and less than or equal to 3, calculating a parameter d which is-0.5677 xq + 2.3962;

if q >3, the parameter d is 0.6931;

(3) calculating the fraction S1 of the first regular model to be 100exp (-d);

the rule model two in the step a23 includes:

(1) calculating the mean and standard deviation of all RR intervals in the available electrocardiogram; then calculating the deviation between each RR interval and the mean value, and judging whether the deviation is greater than the standard deviation; calculating the RR interval ratio r with the deviation exceeding the standard deviation;

(2) inputting the RR interval ratio r, and initializing a parameter d to make d equal to 0;

(3) and (3) judging the r value:

if r is less than or equal to 0.25, the parameter d is 5;

if r is more than 0.25 and less than or equal to 0.35, calculating the parameter d as-26.974 xr + 11.7435;

if r is more than 0.35 and less than or equal to 0.45, calculating the parameter d as-10.896 xr + 6.1477;

if r >0.45, the parameter d is 1.204;

(4) calculating the score S2 of the rule model II to be 100exp (-d);

the rule model three in the step a23 includes:

(1) judging the number n of RR interval groups which accord with complete compensation intermission and approximate premature beat types in the available electrocardiograms: setting four consecutive RR intervals as an RR interval group, comparing the sum of the second RR interval and the third RR interval with the average RR interval, and if the sum of the second RR interval and the third RR interval is less than 2.2 times of the average RR interval and more than 1.1 times of the average RR interval, determining that the second RR interval and the third RR interval conform to a complete compensation interval; determining that the first RR interval is approximately the premature beat type if the first RR interval is greater than the second RR interval and the third RR interval is greater than the fourth RR interval; the RR interval group meeting the two conditions is an RR interval group which accords with complete compensation intermittence and is similar to a premature beat type;

(2) inputting the number n of the RR interval groups which accord with the complete compensation intermission and approximate premature beat type, and initializing a parameter d to make d equal to 0;

(3) and judging the value of n:

if n is less than or equal to 1, the parameter d is 5;

if n is 2, the parameter d is 2.9957;

if n is 3, the parameter d is 1.8971;

if n is 4, the parameter d is 1.204;

if n is more than or equal to 5, the parameter d is 0.6971;

(4) calculating the score S3 of the rule model III to be 100exp (-d);

the rule model four in the step a23 includes:

(1) inputting the height ratio of each waveform P wave R wave in the available electrocardiogram;

(2) if the height ratio of the P wave and the R wave is in the range of 0.1-0.2, determining that the height ratio is a normal height ratio; calculating the ratio P of the height ratio of the R wave of the P wave to all the waveforms in the range in the available electrocardiogram; initializing a parameter d, and enabling d to be 0;

(3) and (3) judging the p value:

if p is less than or equal to 0.4, the parameter d is 5;

if p is more than 0.4 and less than or equal to 0.6, calculating the parameter d as-10.0215 Xp + 9.0086;

if p is more than 0.6 and less than or equal to 0.8, calculating the parameter d as-8.9585 Xp + 8.3708;

if p is more than 0.8 and less than or equal to 0.9, calculating the parameter d as-5.109 Xp + 5.2912;

if p >0.9, the parameter d is 0.6931;

(4) the score of regular model four, S4, is calculated to be 100exp (-d).

In a specific embodiment, the machine learning model judgment in step 2 includes the following steps:

b21, extracting RR interphase, average absolute amplitude, root mean square and wavelet coefficient as features from the input electrocardiogram wave band, decomposing the input heart beat to 4 layers by using db4 wavelet, and taking the wavelet coefficient of a4 frequency band as a feature;

step B22, performing z-score standardization on the features, wherein the z-score standardization is used for enabling the features with different dimensions to have the same scale;

step B23, dimension reduction: for reducing the dimensionality of the features, principal components with weights exceeding 98% are retained;

step B24, inputting the final characteristics obtained after dimensionality reduction into five machine learning submodels which are respectively judged in a linear discriminant analysis model, a Gaussian mixture model, a least square support vector machine, a back propagation neural network model and a probabilistic neural network model to respectively obtain judgment results; the linear discriminant analysis classifies the results into 2 types, atrial fibrillation and non-atrial fibrillation; aiming at the problem of the second classification, the Gaussian mixture model divides a sample into atrial fibrillation and non-atrial fibrillation, and the parameter estimation of the Gaussian mixture model uses an EM algorithm and maximum likelihood estimation; the least squares support vector machine uses radial basis functions as kernel functions for classification; the back propagation neural network comprises a 1-layer input layer, a 1-layer hidden layer and a 1-layer output layer; the neuron number of the input layer is the dimension of the characteristic dimension reduction of the sample; the hidden layer is set to be 6 neurons; the output layer contains 2 neurons, namely atrial fibrillation and non-atrial fibrillation; the probabilistic neural network comprises 1 layer of radial base layer and one layer of competition layer; the radial basal layer comprises two neurons of atrial fibrillation and non-atrial fibrillation; the competition layer determines the class with the maximum probability and assigns 1 to the class;

and step B25, fusing the judgment results to obtain a final result.

In a specific embodiment, each machine learning submodel is trained respectively, model parameters are obtained by training data in advance, and the method is based on the portable equipment only with parameters and directly judges the input waveform.

In a specific embodiment, the results of each submodel in step B25 are fused by voting and proportional scoring, that is, the ratio of atrial fibrillation in the results of each submodel is the final result of the judgment of the machine learning model.

In a specific embodiment, the fusion of the rule model and the machine learning model determination result is as follows: determining whether the two results have larger difference, and if the difference of the scores of the two results is larger than 0.6, the result is invalidated; otherwise, averaging the two values to obtain the final suspected atrial fibrillation degree.

A second aspect of the present invention provides an electrocardiogram atrial fibrillation real-time determination apparatus, including:

the electrocardiogram preprocessing module is used for obtaining an available electrocardiogram marked with waveform characteristics and an electrocardiogram wave band for segmenting the available electrocardiogram; the electrocardiogram preprocessing module comprises: the filtering unit is used for filtering the originally acquired electrocardiogram and removing signal noise to obtain an available electrocardiogram; the waveform identification unit is used for carrying out waveform identification on the available electrocardiogram and identifying and marking P waves and QRS wave groups in the electrocardiogram signals;

the waveform segmentation unit is used for carrying out waveform segmentation on the marked available electrocardiogram and dividing the waveform into a certain wave band which can be used for machine learning to obtain an electrocardiogram wave band;

the rule model is used for inputting the available electrocardiogram marked with the waveform characteristics into the rule model and outputting a rule model result after judgment of the rule model; the rule model includes: the parameter calculation unit is used for calculating RR intervals, P wave heights and R wave heights according to the input available electrocardiograms; an RR interval abnormal value detection unit for detecting and removing abnormal values in RR intervals; the method comprises the following steps that a first rule model for judging the maximum value-to-minimum value ratio of RR intervals, a second rule model for judging the deviation of RR intervals and a mean value, a third rule model for judging the number of RR interval groups which accord with complete compensation gaps and approximate premature beat types, and a fourth rule model for judging the waveform proportion of the height ratio of a P wave R wave in a threshold range are obtained, and the four rule models respectively obtain a first model result S1, a second model result S2, a third model result S3 and a fourth model result S4 through judgment; a judgment unit that judges whether or not S1 is equal to 0: if yes, S is 0; if not, S is S1+ S2-S3-S4; an output unit that outputs the result S;

the machine learning model inputs the segmented electrocardiogram wave bands into the machine learning model, and a machine learning model result is obtained after the judgment of the machine learning model; the machine learning model determination includes: the characteristic extraction unit is used for extracting RR intervals, average absolute amplitudes, root-mean-square and wavelet coefficients as characteristics from the input electrocardiogram wave bands; the normalization unit is used for carrying out z-score normalization processing on the features and enabling the features with different dimensions to have the same scale; a dimension reduction unit: for reducing the dimensionality of the features; the machine learning submodels are respectively a linear discriminant analysis model, a Gaussian mixture model, a least square support vector machine, a back propagation neural network and a probabilistic neural network model, and the characteristics obtained after dimensionality reduction are input into the five machine learning submodels for judgment to respectively obtain judgment results; and the machine learning model calculation unit and each judgment result are subjected to fusion calculation to obtain a final result.

And the calculating unit is used for performing fusion calculation on the rule model result and the machine learning model result to obtain a final result of the doubtful degree of atrial fibrillation.

In a specific embodiment, the rule model one determines a maximum value and a minimum value in all RR intervals, determines an RR interval maximum value and minimum value ratio, determines a range in which the RR interval maximum value and minimum value ratio is located, and calculates a value of a coefficient in response to the range in which the maximum value and minimum value ratio is located, and calculates a score of the model one from the coefficient;

the second rule model determines the mean value and the standard deviation of all RR intervals, determines whether the deviation of each RR interval and the mean value exceeds the standard deviation, determines the number ratio of the RR intervals with the deviation exceeding the standard deviation, determines the range of the number ratio of the RR intervals with the deviation exceeding the standard deviation, responds to the number ratio range, calculates the value of the coefficient, and calculates the fraction of the second model according to the coefficient;

the rule model III processes continuous 4 RR intervals as an RR interval group, determines whether each RR interval group accords with a complete compensation interval, determines whether each RR interval group is similar to a premature beat type, determines the number of the RR interval groups which accord with the complete compensation interval and are similar to the premature beat type, determines the range in which the number of the RR interval groups which accord with the complete compensation interval and are similar to the premature beat type is positioned, responds to the range in which the number is positioned, calculates the value of a coefficient, and calculates the fraction of the model III according to the coefficient;

and the regular model IV determines the height ratio of all the wave forms of the P wave and the R wave, determines whether the height ratio of each wave form of the P wave and the R wave is in a threshold range, determines the waveform proportion of the height ratio of the P wave and the R wave in the threshold range, determines the range of the waveform proportion of the height ratio of the P wave and the R wave in the threshold range, responds to the range of the proportion, calculates the value of the coefficient, and calculates the fraction of the model IV according to the coefficient. A third aspect of the present invention provides an electrocardiogram atrial fibrillation real-time determination system, including:

a memory and one or more processors;

wherein the memory is communicatively coupled to the one or more processors and has stored therein instructions executable by the one or more processors to cause the one or more processors to perform the aforementioned methods.

A fourth aspect of the invention provides a computer-readable storage medium having stored thereon computer-executable instructions operable, when executed by a computing device, to perform the aforementioned method.

In summary, the present invention provides a method, an apparatus, a system and a computer storage medium for real-time judgment of atrial fibrillation of an electrocardiogram, and the method provided by the present invention can analyze the electrocardiogram in real time based on a wearable device. The method extracts features such as RR interphase, root mean square, wavelet coefficient and the like, fully utilizes hidden information in the waveform, and combines rule judgment and a machine learning model to judge atrial fibrillation. And finally, fusing the judgment results of the plurality of models, and giving a score of the atrial fibrillation, namely the doubtful degree of the atrial fibrillation.

The technical scheme of the invention has the following beneficial technical effects:

1. compared with the prior art which only uses the related features of the RR interphase, the method extracts the features of the RR interphase, the root mean square, the wavelet coefficient and the like, fully utilizes hidden information in the waveform, and combines a rule judgment model and a machine learning model to judge atrial fibrillation.

2. Compared with the prior art, the method has the advantages that time delay or long time is judged, only 20s of electrocardiograms are needed to be collected, and atrial fibrillation can be analyzed and judged in real time.

3. Compared with the prior art that the final result gives the absolute judgment of whether the patient has atrial fibrillation, the method gives the possibility that the patient has atrial fibrillation. The decision of the user is facilitated.

Drawings

FIG. 1 is a general flowchart of a method for detecting atrial fibrillation according to an embodiment of the present invention;

FIG. 2 is a flow chart of a rule model in a method for detecting atrial fibrillation according to an embodiment of the present invention;

FIG. 3 is a flow chart of an algorithm of a rule model one of a method for detecting atrial fibrillation according to an embodiment of the present invention;

FIG. 4 is an algorithm flow diagram of a rule model two of a method for detecting atrial fibrillation according to an embodiment of the present invention;

FIG. 5 is an algorithm flow diagram of a rule model three of a method for detecting atrial fibrillation according to an embodiment of the present invention;

FIG. 6 is an algorithm flow diagram of a rule model four of a method for detecting atrial fibrillation according to an embodiment of the present invention;

FIG. 7 is a flow chart of an algorithm of a machine learning model in a method for detecting atrial fibrillation according to an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings in conjunction with the following detailed description. It should be understood that the description is intended to be exemplary only, and is not intended to limit the scope of the present invention. Moreover, in the following description, descriptions of well-known structures and techniques are omitted so as to not unnecessarily obscure the concepts of the present invention.

The invention provides a real-time intelligent judgment method for electrocardiogram atrial fibrillation. As shown in fig. 1.

Fig. 1 is a general flowchart of a method for detecting atrial fibrillation according to an embodiment of the present invention. The original electrocardiogram with the duration of 20s acquired by the portable hardware is filtered to obtain the usable electrocardiogram without interference. Next, waveform identification detection is performed on the available electrocardiogram, and P-wave and QRS-wave complexes are identified. The available electrocardiogram is segmented by the positions of the P wave and the QRS complex. And then inputting the waveforms of the marked feature points and the electrocardiogram wave bands into the rule model and the machine learning model respectively. And finally, fusing the results output by the two models to obtain a final result, namely the doubtful degree of atrial fibrillation.

And the filtering adopts a wavelet threshold method to eliminate noise. The signal is 8-level decomposed using a db6 wavelet. And decomposing the obtained wavelet coefficient, and processing by a soft threshold method to obtain a new wavelet coefficient. And then, signal reconstruction is carried out by the new wavelet coefficient to obtain the filtered electrocardiosignal.

And the waveform identification is used for detecting a QRS complex based on the B-spline biorthogonal wavelet and determining the position of Q, R, S points. In addition, a P-wave is detected based on the first order difference, and a P-point position is determined.

And in the segmentation, the position 0.3s before the R wave crest is taken as a starting point, and the position 0.3s after the R wave crest is taken as an end point, so that the heart beat is obtained. The 6 heartbeats are divided into a band as a sample of the machine learning model.

The rule model includes parameter calculation, abnormal value detection, four rule models, result fusion and the like, as shown in fig. 2.

FIG. 2 is a flow chart of a rule model in a method for detecting atrial fibrillation according to an embodiment of the present invention. In order to judge atrial fibrillation of the 20s electrocardiogram according to a judgment rule, parameters are calculated according to the positions of the P wave and the QRS wave, and a series of RR intervals, P wave heights and R wave heights are obtained. Outliers in the RR intervals are then detected and removed. And then, judging in four rule models respectively according to the characteristics of maximum and minimum value ratios and deviations of RR intervals, the number of RR interval groups which accord with complete compensation intermittence and approximate premature beat types, P wave and P wave height ratios and the like. And finally, fusing the results of the rule models I, II, III and IV to obtain a final result.

And the abnormal value test is used for calculating the mean value of all RR intervals in order to determine whether abnormal values exist in the RR intervals. For each RR interval, a determination is made whether it is less than 1.6 times the mean and greater than 0.5 times the mean. If the threshold range is exceeded, the RR interval is determined to be an outlier and removed.

The rule model I needs to input an RR interval maximum value and minimum value ratio, determines maximum values and minimum values in all RR intervals, determines an RR interval maximum value and minimum value ratio, determines a range of the RR interval maximum value and minimum value ratio, and calculates a coefficient value in response to the range of the maximum value and minimum value ratio, and the coefficient calculates a score of the model I. The rule model one includes 4 layers of judgment and score calculation, as shown in fig. 3.

FIG. 3 is a flow chart of an algorithm of rule model one in a method for detecting atrial fibrillation according to an embodiment of the present invention. The maximum and minimum values of all RR intervals in 20s are input into the model, and the range of the ratio q is determined. If q is less than or equal to 1.1, the parameter d is 5; if q is more than 1.1 and less than or equal to 1.9, calculating the parameter d as-4.745 xq + 10.2195; if q is more than 1.9 and less than or equal to 2.1, the parameter d is 1.204; if q is more than 2.1 and less than or equal to 3, calculating a parameter d which is-0.5677 xq + 2.3962; if q >3, the parameter d is 0.6931. Then, the score of model one, S1, is calculated to be 100exp (-d).

The second rule model needs to input the RR interval ratios with the deviations exceeding the standard deviation, and includes 3-layer judgment and fraction calculation, determining the mean values and the standard deviations of all RR intervals, determining whether the deviation of each RR interval and the mean value exceeds the standard deviation, determining the RR interval number ratios with the deviations exceeding the standard deviation, determining the range where the RR interval number ratios with the deviations exceeding the standard deviation are located, responding to the number ratio range, calculating to obtain the value of the coefficient, and calculating to obtain the fraction of the second model from the coefficient, as shown in fig. 4.

FIG. 4 is a flowchart of an algorithm for rule model two in a method for detecting atrial fibrillation according to an embodiment of the present invention. Before starting, the mean value and standard deviation of all RR intervals within 20s are calculated, then the deviation of each RR interval from the mean value is calculated, and the judgment whether the deviation from the mean value is larger than the standard deviation is made. And inputting the RR interval ratio r with the deviation exceeding the standard deviation into a second model. A determination is made to account for the range of r. If r is less than or equal to 0.25, the parameter d is 5; if r is more than 0.25 and less than or equal to 0.35, calculating the parameter d as-26.974 xr + 11.7435; if r is more than 0.35 and less than or equal to 0.45, calculating the parameter d as-10.896 xr + 6.1477; if r >0.45, the parameter d is 1.204. Then, the score of model two, S2, is calculated to be 100exp (-d).

The rule model three needs to input the number of RR interval groups that meet the full compensation intermission and approximate premature beat type, includes 4-layer judgment and fraction calculation, processes the continuous 4 RR intervals as one RR interval group, determines whether each RR interval group meets the full compensation intermission, determines whether each RR interval group is approximate to the premature beat type, determines the number of RR interval groups that meet the full compensation intermission and approximate premature beat type, determines the range in which the number of RR interval groups that meet the full compensation intermission and approximate premature beat type is located, and calculates the value of the coefficient in response to the range in which the number is located, and calculates the fraction of the model three from the coefficient, as shown in fig. 5.

FIG. 5 is a flow chart of an algorithm of rule model three in a method for detecting atrial fibrillation according to an embodiment of the present invention. Four consecutive RR intervals are treated as a group of RR intervals. To determine whether a group of RR intervals corresponds to a fully compensated pause, the sum of the second RR interval and the third RR interval is compared to the average RR interval. Determining that the second RR interval and the third RR interval conform to a full compensation intermission if the sum of the second RR interval and the third RR interval is less than 2.2 times the average RR interval and greater than 1.1 times the average RR interval. Next, to determine whether it satisfies the approximate premature beat type, the four RR intervals are compared. If the first RR interval is greater than the second RR interval and the third RR interval is greater than the fourth RR interval, determining that it is approximately the premature beat type. Inputting the number n of RR interval groups which are in line with complete compensation intermittence and approximate premature beat types into a third model. A determination is made of a range of numbers n. If n is less than or equal to 1, the parameter d is 5; if n is 2, the parameter d is 2.9957; if n is 3, the parameter d is 1.8971; if n is 4, the parameter d is 1.204; if n is greater than or equal to 5, the parameter d is 0.6971. The score for model three, S3, is then calculated to be 100exp (-d).

And the regular model IV needs to input the height ratio of each waveform P wave R wave, and comprises the steps of calculating the number and the occupation ratio of the waveforms with the PR height ratio within the threshold range, judging by 4 layers and calculating the fraction, determining the height ratio of all the waveforms P wave R waves, determining whether the height ratio of each waveform P wave R wave is within the threshold range, determining the occupation ratio of the P wave R wave height ratio within the threshold range, determining the range of the waveform occupation ratio of the P wave R wave height ratio within the threshold range, responding to the range of the occupation ratio, calculating to obtain the value of the coefficient, and calculating to obtain the fraction of the model IV by the coefficient. As shown in fig. 6.

FIG. 6 is a flow chart of an algorithm of rule model four in a method for detecting atrial fibrillation according to an embodiment of the present invention. The height ratio of the R wave of each wave P wave is input into the model four. To determine whether the P-wave R-wave height ratio is a normal P-wave R-wave height ratio, a determination is made whether the P-wave R-wave height ratio is within a threshold range. If the P-wave R-wave height ratio is in the range of 0.1-0.2, the P-wave R-wave height ratio is determined to be the normal height ratio. The ratio of the height of the P-wave R-wave to the height of the wave within the threshold range to all the waves within 20s is then calculated. A determination is made that the waveform in the normal range is in proportion to the p-range. If p is less than or equal to 0.4, the parameter d is 5; if p is more than 0.4 and less than or equal to 0.6, calculating the parameter d as-10.0215 Xp + 9.0086; if p is more than 0.6 and less than or equal to 0.8, calculating the parameter d as-8.9585 Xp + 8.3708; if p is more than 0.8 and less than or equal to 0.9, calculating the parameter d as-5.109 Xp + 5.2912; if p >0.9, the parameter d is 0.6931. Then, the score of model four, S4, is calculated to be 100exp (-d).

The results are merged by first making a determination of whether the value of model one result S1 is 0. If S1 is equal to 0, the final score S is equal to 0; if S1 ≠ 0, the final score S ═ S1+ S2-S3-S4 is calculated. This is the final score of the rule model.

The machine learning model in fig. 1 includes feature extraction, normalization, dimension reduction, 5 machine learning models, and model fusion, as shown in fig. 7.

FIG. 7 is a flow chart of an algorithm for a machine learning model in a method for detecting atrial fibrillation according to an embodiment of the present invention. Firstly, carrying out feature extraction on the divided electrocardiogram segments with 6 heartbeats as units, wherein the features are adopted as follows: RR interval, mean absolute amplitude, root mean square, wavelet coefficients. After the features are normalized, principal component analysis is used to reduce dimensions. And then inputting the samples into a linear discriminant analysis model, a Gaussian mixture model, a least square support vector machine, a back propagation neural network and a probabilistic neural network model to respectively obtain classification results. And finally, fusing the results of all models to obtain a final result.

And the wavelet coefficient characteristics are that 4-layer decomposition is carried out on the samples by using db4 wavelets, and the wavelet coefficients of a4 frequency band are taken as characteristics.

And (4) performing principal component analysis, and keeping the first few principal components with the weight exceeding 98% to finish dimensionality reduction.

And the linear discriminant analysis divides the result into 2 types, namely atrial fibrillation and non-atrial fibrillation.

The Gaussian mixture model also divides the sample into atrial fibrillation and non-atrial fibrillation according to the problem of the second classification. The parameter estimation uses the EM algorithm and the maximum likelihood estimation.

The least squares support vector machine uses radial basis functions as kernel functions for classification.

The back propagation neural network comprises a 1-layer input layer, a 1-layer hidden layer and a 1-layer output layer. The neuron number of the input layer is the dimension of the characteristic dimension reduction of the sample; the hidden layer is set to be 6 neurons; the output layer contains 2 neurons, atrial fibrillation and non-atrial fibrillation.

The probabilistic neural network comprises 1 radial basic layer and one competition layer. The radial basal layer contains two neurons, atrial fibrillation and non-atrial fibrillation. The contention layer determines the class with the highest probability and assigns it a value of 1.

The 5 machine learning models are trained in advance by using data marked by doctors in hospitals to obtain model parameters. The input waveform is directly judged during analysis.

And the models are fused, and when a sample is input into each model to obtain a result, all the models vote together to give respective results. Wherein the ratio of the atrial fibrillation is the atrial fibrillation judgment score of the machine learning model.

Finally, the results of the rule model and the machine learning model are integrated. To determine if there is a large difference in the results of the two models, the two models are compared. If the difference of the scores of the two is more than 0.6, the result is invalidated. Otherwise, averaging the two values to obtain the final atrial fibrillation doubtful degree.

Another aspect of the present invention provides an apparatus for real-time determining atrial fibrillation with an electrocardiogram, comprising:

the electrocardiogram preprocessing module is used for obtaining an available electrocardiogram marked with waveform characteristics and an electrocardiogram wave band for segmenting the available electrocardiogram;

the rule model is used for inputting the available electrocardiogram marked with the waveform characteristics into the rule model and outputting a rule model result after judgment of the rule model;

the machine learning model inputs the segmented electrocardiogram wave bands into the machine learning model, and a machine learning model result is obtained after the judgment of the machine learning model;

Further, the electrocardiogram preprocessing module comprises:

the filtering unit is used for filtering the originally acquired electrocardiogram and removing signal noise to obtain an available electrocardiogram;

the waveform identification unit is used for carrying out waveform identification on the available electrocardiogram and identifying and marking P waves and QRS wave groups in the electrocardiogram signals;

and the waveform segmenting unit is used for segmenting the waveforms of the marked available electrocardiograms and dividing the waveforms into certain wave bands which can be used for machine learning to obtain the electrocardiogram wave bands.

Further, the rule model includes:

the parameter calculation unit is used for calculating RR intervals, P wave heights and R wave heights according to the input available electrocardiograms;

an RR interval abnormal value detection unit for detecting and removing abnormal values in RR intervals;

the method comprises the following steps that a first rule model for judging the maximum value-to-minimum value ratio of RR intervals, a second rule model for judging the deviation of RR intervals and a mean value, a third rule model for judging the number of RR interval groups which accord with complete compensation gaps and approximate premature beat types, and a fourth rule model for judging the waveform proportion of the height ratio of a P wave R wave in a threshold range are obtained, and the four rule models respectively obtain a first model result S1, a second model result S2, a third model result S3 and a fourth model result S4 through judgment;

a judgment unit that judges whether or not S1 is equal to 0: if yes, S is 0; if not, S is S1+ S2-S3-S4;

and an output unit which outputs the result S.

Further, the machine learning model determining comprises:

the characteristic extraction unit is used for extracting RR intervals, average absolute amplitudes, root-mean-square and wavelet coefficients as characteristics from the input electrocardiogram wave bands;

and the normalization unit is used for normalizing the features and enabling the features with different dimensions to have the same dimension. Here z-score normalization is used: y ═ standard deviation of (average of X-X)/X; x is the set of all values of the feature, and X is a single value.

A dimension reduction unit: for reducing the dimensionality of the features;

the machine learning submodels are respectively a linear discriminant analysis model, a Gaussian mixture model, a least square support vector machine, a back propagation neural network and a probabilistic neural network model, and the final characteristics obtained after dimensionality reduction are input into the five machine learning submodels for judgment to respectively obtain judgment results;

and the calculation unit and each judgment result are subjected to fusion calculation to obtain a final result.

Yet another aspect of the present invention provides an electrocardiogram atrial fibrillation real-time judging system, which includes:

a memory and one or more processors;

A final aspect of the invention provides a computer-readable storage medium having stored thereon computer-executable instructions operable, when executed by a computing device, to perform the aforementioned method.

It is to be understood that the above-described embodiments of the present invention are merely illustrative of or explaining the principles of the invention and are not to be construed as limiting the invention. Therefore, any modification, equivalent replacement, improvement and the like made without departing from the spirit and scope of the present invention should be included in the protection scope of the present invention. Further, it is intended that the appended claims cover all such variations and modifications as fall within the scope and boundaries of the appended claims or the equivalents of such scope and boundaries.

Claims

1. A real-time judgment method for atrial fibrillation with electrocardiogram is characterized by comprising the following steps:

step 2, inputting the available electrocardiogram marked with the waveform characteristics into a rule model, and outputting a rule model result after the rule model is judged based on a ventricular escape medical judgment rule; inputting the segmented electrocardiogram wave bands into a machine learning model trained through hospital data, and obtaining a machine learning model result after the machine learning model is judged, wherein the rule model judgment comprises the following steps:

step A23, obtaining a first model result S1, a second model result S2, a third model result S3 and a fourth model result S4 respectively through judgment of a first rule model for judging the maximum value and minimum value ratio of RR intervals, a second rule model for judging the deviation of RR intervals and a mean value, a third rule model for judging the number of RR interval groups which accord with complete compensation gaps and are premature beat types, and a fourth rule model for judging the waveform ratio of the height ratio of the P wave R wave in a threshold range;

step A24, determine S1 equal to 0: if yes, S is 0; if not, S is S1+ S2-S3-S4; s is an output result;

step A25, outputting S;

2. The method for determining atrial fibrillation according to claim 1, wherein the abnormal value detection in step a22 is used to determine whether an abnormal value exists in RR intervals, and first, the average value of all RR intervals is calculated; and judging whether the value of each RR interval is more than 0.5 time of the mean value and less than 1.6 times of the mean value, if not, exceeding the threshold range, determining that the RR interval is an abnormal value and removing the abnormal value.

3. The method for real-time atrial fibrillation determination according to claim 1,

the rule model one in the step a23 includes:

(2) and (3) judging the q value:

if q is less than or equal to 1.1, the parameter d is 5;

if q is more than 1.1 and less than or equal to 1.9, the parameter d is-4.745 xq + 10.2195;

if q is more than 1.9 and less than or equal to 2.1, the parameter d is 1.204;

if q is more than 2.1 and less than or equal to 3, the parameter d is-0.5677 xq + 2.3962;

if q >3, the parameter d is 0.6931;

(3) calculating the fraction S1 of the first regular model to be 100exp (-d);

the rule model two in the step a23 includes:

(3) and (3) judging the r value:

if r is less than or equal to 0.25, the parameter d is 5;

if r is more than 0.25 and less than or equal to 0.35, the parameter d is-26.974 xr + 11.7435;

if r is more than 0.35 and less than or equal to 0.45, the parameter d is-10.896 xr + 6.1477;

if r >0.45, the parameter d is 1.204;

(4) calculating the score S2 of the rule model II to be 100exp (-d);

the rule model three in the step a23 includes:

(1) judging the number n of RR interval groups which accord with complete compensation intermission and are of a premature beat type in the available electrocardiograms: setting four consecutive RR intervals as an RR interval group, comparing the sum of the second RR interval and the third RR interval with the average RR interval, and if the sum of the second RR interval and the third RR interval is less than 2.2 times of the average RR interval and more than 1.1 times of the average RR interval, determining that the second RR interval and the third RR interval conform to a complete compensation interval; determining that the first RR interval is a premature beat type if the first RR interval is greater than the second RR interval and the third RR interval is greater than the fourth RR interval; the RR interval group meeting the two conditions simultaneously is an RR interval group which accords with complete compensation intermission and is of a premature beat type;

(2) inputting the number n of the RR interval groups which accord with the complete compensation intermission and are in the premature beat type, and initializing a parameter d to enable d to be 0;

(3) and judging the value of n:

if n is less than or equal to 1, the parameter d is 5;

if n is 2, the parameter d is 2.9957;

if n is 3, the parameter d is 1.8971;

if n is 4, the parameter d is 1.204;

if n is more than or equal to 5, the parameter d is 0.6971;

(4) calculating the score S3 of the rule model III to be 100exp (-d);

the rule model four in the step a23 includes:

(3) and (3) judging the p value:

if p is less than or equal to 0.4, the parameter d is 5;

if p is more than 0.4 and less than or equal to 0.6, the parameter d is-10.0215 Xp + 9.0086;

if p is more than 0.6 and less than or equal to 0.8, the parameter d is-8.9585 Xp + 8.3708;

if p is more than 0.8 and less than or equal to 0.9, the parameter d is-5.109 Xp + 5.2912;

if p >0.9, the parameter d is 0.6931;

(4) the score of regular model four, S4, is calculated to be 100exp (-d).

4. The method for real-time judgment of atrial fibrillation according to the electrocardiogram of claim 1, wherein the machine learning model judgment in the step 2 comprises the following steps:

and step B25, fusing the judgment results to obtain a final result.

5. The method according to claim 4, wherein each machine learning submodel is trained separately, and the model parameters are obtained by training data in advance, and the method is based on the portable device with only parameters stored therein, and the input waveform is directly judged.

6. The method as claimed in claim 4, wherein the results of each submodel in step B25 are fused by voting and proportional scoring, that is, the ratio of atrial fibrillation in the results of each submodel is the final result of the machine learning model.

7. The method according to any of claims 1-4, wherein the rule model and the machine learning model are fused to determine the atrial fibrillation: determining whether the two results are different, and if the difference of the scores of the two results is greater than 0.6, the result is invalidated; otherwise, averaging the two values to obtain the final suspected atrial fibrillation degree.

8. An electrocardiogram atrial fibrillation real-time judging device, comprising:

the waveform segmentation unit is used for carrying out waveform segmentation on the marked available electrocardiogram and dividing the waveform into wave bands which can be used for machine learning to obtain electrocardiogram wave bands;

the rule model is used for inputting the available electrocardiogram marked with the waveform characteristics into the rule model and outputting a rule model result after judgment of the rule model; the rule model includes: the parameter calculation unit is used for calculating RR intervals, P wave heights and R wave heights according to the input available electrocardiograms; an RR interval abnormal value detection unit for detecting and removing abnormal values in RR intervals; the method comprises the following steps that a first rule model for judging the maximum value-to-minimum value ratio of RR intervals, a second rule model for judging the deviation of RR intervals and a mean value, a third rule model for judging the number of RR interval groups which accord with complete compensation gaps and are premature beat types, and a fourth rule model for judging the waveform proportion of the height ratio of a P wave R wave in a threshold range, wherein the four rule models respectively obtain a first model result S1, a second model result S2, a third model result S3 and a fourth model result S4 through judgment; a judgment unit that judges whether or not S1 is equal to 0: if not, S is S1+ S2-S3-S4; if yes, S is 0; s is an output result; an output unit that outputs the result S;

the machine learning model inputs the segmented electrocardiogram wave bands into the machine learning model, and a machine learning model result is obtained after the judgment of the machine learning model; the machine learning model determination includes: the characteristic extraction unit is used for extracting RR intervals, average absolute amplitudes, root-mean-square and wavelet coefficients as characteristics from the input electrocardiogram wave bands; the normalization unit is used for carrying out z-score normalization processing on the features and enabling the features with different dimensions to have the same scale; a dimension reduction unit: for reducing the dimensionality of the features; the machine learning submodels are respectively a linear discriminant analysis model, a Gaussian mixture model, a least square support vector machine, a back propagation neural network model and a probabilistic neural network model, and the characteristics obtained after dimensionality reduction are input into the five machine learning submodels for judgment to respectively obtain judgment results; the machine learning model calculation unit performs fusion calculation on each judgment result to obtain a final result;

9. The atrial fibrillation real-time determining apparatus of claim 8, wherein,

the rule model I determines maximum values and minimum values in all RR intervals, determines the ratio of the maximum values and the minimum values of the RR intervals, determines the range of the ratio of the maximum values and the minimum values of the RR intervals, calculates the value of a coefficient in response to the range of the ratio of the maximum values and the minimum values, and calculates the score of the model I according to the coefficient;

the second rule model determines the mean value and the standard deviation of all RR intervals, determines whether the deviation of each RR interval and the mean value exceeds the standard deviation, determines the number ratio of the RR intervals with the deviation exceeding the standard deviation, determines the range of the number ratio of the RR intervals with the deviation exceeding the standard deviation, and calculates the value of a coefficient in response to the range of the number ratio, and calculates the fraction of the second model according to the coefficient;

the rule model III processes continuous 4 RR intervals as an RR interval group, determines whether each RR interval group accords with a complete compensation interval, determines whether each RR interval group is a premature beat type, determines the number of the RR interval groups which accord with the complete compensation interval and are the premature beat type, determines the range in which the number of the RR interval groups which accord with the complete compensation interval and are the premature beat type is positioned, responds to the range in which the number is positioned, calculates the value of the coefficient, and calculates the fraction of the model III according to the coefficient;

and the regular model IV determines the height ratio of all the wave forms of the P wave and the R wave, determines whether the height ratio of each wave form of the P wave and the R wave is in a threshold range, determines the waveform proportion of the height ratio of the P wave and the R wave in the threshold range, determines the range of the waveform proportion of the height ratio of the P wave and the R wave in the threshold range, responds to the range of the proportion, calculates the value of the coefficient, and calculates the fraction of the model IV according to the coefficient.

10. An electrocardiogram atrial fibrillation real-time judging system, which is characterized by comprising:

a memory and one or more processors;

wherein the memory is communicatively coupled to the one or more processors and has stored therein instructions executable by the one or more processors to cause the one or more processors to perform the method of any of claims 1-7.

11. A computer-readable storage medium having stored thereon computer-executable instructions operable, when executed by a computing device, to perform the method of any of claims 1-7.