CN114176519A

CN114176519A - Non-contact electrocardiosignal quality classification method

Info

Publication number: CN114176519A
Application number: CN202111496510.9A
Authority: CN
Inventors: 列智坤; 陈炜; 吴咏霖; 祝国强; 崔睿; 陈晨
Original assignee: Fudan University
Current assignee: Fudan University
Priority date: 2021-12-09
Filing date: 2021-12-09
Publication date: 2022-03-15

Abstract

The invention belongs to the field of health detection, and particularly relates to a non-contact electrocardiosignal quality classification method. The wavelet median threshold method is used for removing the baseline drift and the myoelectricity interference noise of the signals, and a fusion classifier model with strong robustness and stability is adopted for carrying out high-precision classification on the quality of the electrocardiosignals, so that the utilization rate of the signals is improved, and the pathological analysis of a patient can be further accurately carried out. The method is suitable for being applied to a non-contact electrocardiosignal detection system for long-time day and night detection, and solves the problems of low efficiency of dynamic electrocardiosignal feature extraction, signal noise and obvious motion artifact. And the complexity of the algorithm and the space of the model are low, and the time for predicting and classifying is reduced, so that the family application popularization of portable intelligent medical treatment is realized.

Description

Non-contact electrocardiosignal quality classification method

Technical Field

The invention belongs to the field of health detection, and relates to a quality classification method based on non-contact electrocardiosignals.

Background

Cardiovascular diseases have become a leading cause of human death, and their morbidity and mortality rates have remained high, especially in areas of medical resource scarcity. The average number of deaths caused by heart diseases in China exceeds 540000 every year, cardiovascular diseases are long-term accumulated results, and the disease deterioration can be avoided as far as possible only by early discovery. Electrocardiographic monitoring is an indispensable technology in clinical and medical care, and has important significance for monitoring physical signs, diagnosing or early warning cardiovascular diseases and guiding clinical decisions [1 ]. However, the traditional electrocardio monitoring device has defects in convenience and comfort, and the traditional commonly used method for measuring the electrocardio uses a wet electrode method, firstly needs to clean the skin, removes thick hair and skin cuticle, then wipes alcohol or coats conductive gel, and then attaches the wet electrode on the body of a patient to carry out the electrocardio measurement [2 ]. Obviously, the traditional measurement method is not convenient, especially for infants and people with sensitive skin, and is easy to cause panic and nervous emotion of patients, so that the electrocardio measurement result is influenced, and the traditional electrocardio measurement method is inconvenient for long-time monitoring and night monitoring of the patients.

With the continuous improvement of instrument performance and detection technology, more and more researchers begin to research the detection method of non-contact electrocardiosignal, and the development of portable, wearable and non-inductive electrocardio monitoring devices is promoted by the breakthrough of microelectronic technology and integrated circuit technology [3 ]. However, the novel electrocardiograph measurement device overcomes the defects of the conventional device and is accompanied by some unavoidable defects, such as the problems that the signal is easily interfered by environmental noise, obvious motion artifacts exist, the waveform signal is not obvious, and the like, so that it is necessary to design an algorithm model to classify the quality of the electrocardiograph signal acquired in a non-contact manner so as to improve the processing efficiency and the utilization rate of the signal. At present, there are many algorithms for machine learning in the aspect of classification of cardiac electrical signal quality, such as training a support vector machine SVM to classify by extracting a plurality of feature values of an ECG signal [4], classifying by using a high-dimensional feature of the ECG and using a KNN model based on euclidean distance metric [5], extracting several non-reference features from the ECG signal and then using a binary decision tree model to classify the signal [6], performing electrocardiographic classification using a random forest model based on time-domain and frequency-domain features [7], and because of different classification standards of different classifiers, the classification accuracy is related to a plurality of factors such as statistical distribution features of classified data, sizes of training data samples, and structures of the classifiers themselves. In the last decade, a great deal of research is directed to detecting the electrical activity information of the heart by a non-invasive and non-interference sensing method, and the method is expected to overcome the limitation of the traditional electrocardiogram monitoring and ensure the information richness, observation interpretability and measurement stability of signals.

Reference documents:

[1] E. J. Benjamin, S. S. Virani, C. W. Callaway, A. M. Chamberlain, A. R. Chang, S. Cheng, S. E. Chiuve, M. Cushman, F. N. Delling, R. Deo et al., “Heart disease and stroke statistics-2018 update: a report from the american heart association.” Circulation, vol. 137, no. 12, pp.67–97, 2018.

[2] S. Peng, S. Bao and W. Chen, "Capacitive Coupled Electrodes based Non-contact ECG Measurement System with Real-time Wavelet Denoising Algorithm," 2019 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2019, pp. 6587-6590, doi: 10.1109/EMBC.2019.8856885.

[3] Y. M. Chi and G. Cauwenberghs, "Wireless Non-contact EEG/ECG Electrodes for Body Sensor Networks," 2010 International Conference on Body Sensor Networks, 2010, pp. 297-301, doi: 10.1109/BSN.2010.52.

[4] Z. Ge, Z. Zhu, P. Feng, S. Zhang, J. Wang and B. Zhou, "ECG-Signal Classification Using SVM with Multi-feature," 2019 8th International Symposium on Next Generation Electronics (ISNE), 2019, pp. 1-3, doi: 10.1109/ISNE.2019.8896430.

[5] E. Prabhakararao and S. Dandapat, "Automatic Quality Estimation of 12-lead ECG for Remote Healthcare Monitoring Systems," 2018 IEEE-EMBS Conference on Biomedical Engineering and Sciences (IECBES), 2018, pp. 554-559, doi: 10.1109/IECBES.2018.8626686.

[6] B. U. Demirel and Y. Serinağaoğlu, "Quality Assessment of ECG Signals Based on Support Vector Machines and Binary Decision Trees," 2020 28th Signal Processing and Communications Applications Conference (SIU), 2020, pp. 1-4, doi: 10.1109/SIU49456.2020.9302262.

[7] M. Kropf, D. Hayn and G. Schreier, "ECG classification based on time and frequency domain features using random forests," 2017 Computing in Cardiology (CinC), 2017, pp. 1-4, doi。

disclosure of Invention

The invention aims to develop on the basis of the prior art, provides a quality classification algorithm for non-contact acquisition of electrocardiosignals based on a capacitive coupling electrode principle, is suitable for being applied to a non-contact electrocardiosignal detection system for long-time day and night detection, and solves the problems of low efficiency of dynamic electrocardiosignal feature extraction, signal noise and obvious motion artifact. The fusion classifier model with strong robustness and stability is adopted to carry out high-precision classification on the quality of the electrocardiosignals, so that the utilization rate of the signals is improved, and the pathological analysis of a patient is further accurately carried out. The technical scheme of the invention is specifically introduced as follows.

In order to achieve the purpose of the invention, the invention provides a non-contact electrocardiosignal quality classification method, in particular to

The method comprises the following steps:

step 1, collecting electrocardiosignals by using non-contact electrocardio collecting equipment;

step 2, after filtering and denoising the collected electrocardiosignals, carrying out segmentation processing and labeling on the electrocardiosignals, and classifying the electrocardiosignal quality according to the purpose;

step 3, extracting local features of the filtered and denoised electrocardiosignals, wherein the local features comprise a kurtosis value K, a skewness value S, a signal range F1, a signal standard deviation F2, an average R-R interval F3 and various features of the R wave number F4 in a time domain, a frequency domain and a nonlinear domain in a sampling period;

step 4, standardizing various mathematical characteristics of the signal waveform obtained in the step 3, calculating an average mean and a standard deviation std of the characteristics, and establishing a characteristic matrix F = [ K, S, F1, F2, F3 and F4 … … ], wherein each group of data has a data label marked by an expert and corresponds to the data label;

step 5, dividing the electrocardiosignals into a training set and a testing set, inputting various characteristics of signal waveforms and signal quality categories into a fusion classifier in the training set to realize the training of a fusion model, then inputting sample signals in the testing set into the fusion classifier for classification, and evaluating and verifying the stability and robustness of the model through the testing set; the method specifically comprises the following steps:

(1) building and training a plurality of classifier models including a support vector machine, a K neighbor model and a decision tree model;

(2) fusing a plurality of classifier models through a voting mechanism to form a fusion model, determining the hyper-parameters of various classifiers by using a cross validation method, and realizing the training of the fusion model; the voting mechanism is as follows: when most classifiers evaluate the same signal to be in a certain class, the signal is evaluated to be in the class, and when the classification results of a plurality of classifiers are inconsistent, the classification result of one classifier is randomly selected;

and 6, inputting the electrocardiosignals to be classified into the fusion model trained in the step 5 after the electrocardiosignals are processed according to the steps 2 to 4, and obtaining the electrocardiosignal quality classification result.

In the invention, in step 2, a wavelet median threshold method is adopted to filter and denoise the electrocardiosignals; wherein the mathematical expression of the threshold function is:

wherein log is logarithmized, noise variance, and has the value: (ii) a Wherein abs is absolute value removal, mean is median, x is a signal vector, and n is the length of a wavelet coefficient vector; l is the wavelet coefficient decomposition layer number and is determined to be 4; m is the current layer, the value range is {2,3,4}, and the wavelet function is determined as DB 1.

In the invention, in the step 2, the signals are classified into A, B, C types, wherein, A type is a clear ECG signal which can be directly used clinically, the signal base line is stable due to regular R wave and T wave, and the ECG rhythm analysis and the morphology analysis can be carried out; the B type is a signal which has an R wave but has unclear integral signal and can extract ECG waveform characteristics after being processed, is an irregular R wave and a T wave and is generally only used for analyzing the electrocardio rhythm; the class C is a clinically unavailable signal with obvious baseline drift and large noise, the signal-to-noise ratio is extremely low, and simultaneously, step signals, saturation phenomena and the like occur, so that the QRS wave complex can hardly be judged.

In the invention, in step 3:

(1) and extracting kurtosis and skewness of the signal according to the signal waveform, wherein the kurtosis is a statistic for describing the steep degree of all values in the distribution form in the population. The calculation formula is as follows: the skewness is a measure of the skew direction and degree of the statistical data distribution, and the calculation formula is as follows: wherein K represents kurtosis; a value representing the ith sample, representing the average value, n being the number of samples;

(2) calculating the range, standard deviation, average R-R interval and the number of R waves in a period of the sample signal according to the signal waveform;

the range is used for counting the difference between the maximum value and the minimum value in the data, and F1 represents the value; the standard deviation reflects the degree of dispersion of a data set, and is a measure of the degree of dispersion of the mean values of a set of data, and is calculated as follows: (ii) a The average RR interval is the average of the time intervals between R-waves of all QRS waveforms in the sampled signal, the value of which is represented by F3; the number of R waves refers to the total number of R waves in the sampled signal, and is calculated based on the QRS waveform detection algorithm (Pan and Tompkins (P & T) and 'wqrs' algorithm), and the value is represented by F4.

In the invention, in step 4, a confusion matrix of the model is calculated through the test set, and the trained model is evaluated by using Recall Ratio (Recall), Positive Predictive Ratio (Positive Predictive Ratio), Accuracy (Accuracy) and F1-Score, so as to measure the stability and robustness of the model.

Compared with the prior art, the invention has the beneficial effects that:

compared with other artificial intelligent electrocardio quality classification methods, the method is more suitable for being applied to a non-contact electrocardio signal detection device for long-time day and night detection, is favorable for solving the problems of low dynamic electrocardio signal characteristic extraction, signal noise and obvious motion artifact, and the test set and the training set of the model are both composed of electrocardio of different patients, thereby greatly overcoming the problem of personal electrocardio specificity. And the complexity of the algorithm and the space of the model are low, the prediction classification time is shortened, and the quasi-real-time effect is achieved, so that the family application and popularization of portable intelligent medical treatment are realized.

Drawings

FIG. 1 is a flow chart of the present invention.

FIG. 2 is a fusion classifier model classification process.

Fig. 3 is a waveform of a class a signal.

Fig. 4 shows a waveform of a B-type signal.

Fig. 5 shows a waveform of a class C signal.

FIG. 6 is a resulting confusion matrix for a fused classifier model.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described below with reference to the accompanying drawings. It should be understood, however, that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

Example 1

Electrocardiography is widely recognized as the gold standard for cardiac monitoring, and many cardiovascular diseases can be better diagnosed, controlled and prevented by continuously monitoring and analyzing cardiac-related physiological signals. Aiming at the electrocardiosignals acquired by the non-contact electrocardiosignal monitoring method, in order to overcome the problems of environmental noise interference, obvious motion artifacts, unobvious waveforms and the like, the electrocardiosignals are filtered and then the characteristics of the electrocardiosignals are extracted, and then the acquired electrocardiosignals are classified into three categories of A/B/C by a classifier and then are subjected to clinical pathological analysis.

The flow chart of the invention for classifying non-contact electrocardiosignals is shown in figure 1, and comprises the following steps:

step 1, collecting electrocardiosignals collected by a non-contact electrocardio monitoring device.

And 2, removing noises such as baseline drift, electromyographic interference and the like from the acquired electrocardiosignals by using a wavelet median threshold method. Then, the collected electrocardiosignals are segmented at intervals of 5 seconds, effective signal data with obvious characteristics are selected for 2000 cases, the extracted electrocardiosignals are manually labeled, and the signal quality is divided into A, B, C categories according to the purpose, as shown in fig. 3-5: a type: clear ECG signals (available for direct clinical use); b type: r-waves exist but the overall signal is ambiguous (processed to extract ECG waveform features); class C: there is significant baseline drift and noisy signals (such signals are not clinically useful).

And 3, extracting local features of the filtered and denoised electrocardiosignals, wherein the local features comprise various features of a kurtosis value, a deviation value, a signal range, a signal standard deviation, an average R-R interval, the number of R waves in a sampling period and the like in a time domain, a frequency domain and a nonlinear domain.

3.1 carry on QRS complex detection to the electrocardiosignal with Pan-Tompkins method, differentiate to obtain QRS slope information first to the signal, then ask square to strengthen the frequency response curve slope of the derivative, limit the false positive caused by T wave higher than the general spectral energy, and then use the moving window integral to produce the signal about slope and width information of QRS wave, confirm the position of R wave according to the adaptive threshold finally, the updating mode of the adaptive threshold THREADOLD 1 is: SPK1=0.125pEAk1+0.875SPK1 '; NPK1=0.125pEAk1+0.875NPK 1'; threshold 1= NPK1+0.25(SPK1-NPK1);

where SPK1 is a normal PEAK, NPK1 is a noise PEAK, PEAK1 represents an overall PEAK, which is considered a normal PEAK if a signal PEAK greater than the threshold value threshold 1 is detected, and a noise PEAK otherwise.

3.2 extracting kurtosis and skewness of the signal according to the signal waveform.

And 3.3, calculating the RR interval characteristics and the number of R waves in the segmented signals by taking the R waves as a reference, wherein the RR interval characteristics comprise a pre-RR interval, a current RR interval and a post-RR interval. The method for calculating the pre-RR interval comprises the following steps: preR = Xn-Xn-1; xn-1 is the position of the previous R-wave, Xn is the position of the current R-wave, and the post RR interval is calculated by postR = Xn + 1-Xn; xn +1 is the position of the next R-wave, and the current RR interval is the average of the first 10R-waves: localR = (Xn-10 + Xn-9 … Xn)/10; and finally, calculating the range error and the standard deviation of the sample signal according to the signal waveform.

And 4, standardizing various mathematical characteristics of the signal waveform obtained in the step 3, calculating the mean value mean and standard deviation std of the characteristics, establishing a characteristic matrix F = [ K, S, F1, F2, F3 and F4 … … ], wherein each group of data has a data label marked by an expert corresponding to the data label, and training a classifier model by following a characteristic engineering combined label mode.

And 5, designing a fusion classifier and a voting mechanism, and calculating an evaluation index of the model.

5.1 designing fusion classifier models of a plurality of machine learning methods such as a support vector machine, a K neighbor model, a decision tree model and the like to form an integral classification system, and designing a voting mechanism of the fusion classifier: when most classifiers rate the same signal to a certain grade, we have enough reason to rate the signal to the grade, and when the classification result of each classifier is inconsistent, we randomly select the classification result of one classifier. After multiple training and cross validation, the signal is finally classified into A, B, C grades.

5.2 the confusion matrix for the computational model is shown in FIG. 6.

The stability and robustness of the model were measured by evaluating the trained model using Recall (Recall), Positive Predictive Ratio (Positive Predictive Ratio), Accuracy (Accuracy) and F1-Score, and the results are shown in table 1. Training a fusion classifier model according to the processed training set, evaluating the accuracy of the model through a test set, and comparing with the classification result of a single classifier, as shown in table 2.

TABLE 1 model Performance indices of fusion classifier

Type (B)	Acc	P₊	Se	F-Score
					A	98.26%	98.26%	98.26%	98.26%
B	97.81%	98.35%	97.81%	98.08%
					C	98.89%	98.34%	98.89%	98.61%
Mean value of	98.31%	98.32%	98.32%	98.25%

TABLE 2 fusion of classifier and common model classification results

	SVM	KNN	Decision Tree	Fusion classifier
					Accuracy of	88.9%	84.8%	87.1%	98.3%

The result shows that the fusion model has stability and robustness, and the result of the fusion model is more accurate compared with that of a single classification model.

Claims

1. A non-contact electrocardiosignal quality classification method is characterized by comprising the following specific steps:

step 5, dividing the electrocardiosignals into a training set and a testing set, inputting various characteristics of signal waveforms and signal quality categories into a fusion classifier in the training set to realize the training of the fusion model, then inputting sample signals in the testing set into the fusion classifier for classification, and evaluating and verifying the stability and robustness of the fusion model through the testing set; the method specifically comprises the following steps:

2. The method for classifying the quality of the electrocardiosignals according to claim 1, wherein in the step 2, the electrocardiosignals are filtered and denoised by adopting a wavelet median threshold method.

3. The method for classifying the quality of electrocardiosignals according to claim 1, wherein in step 2, the signals are classified into A, B, C types, wherein A type is a clear ECG signal which can be directly used clinically, B type is a signal which has R wave but has not clear overall signal to be processed so as to extract the characteristics of ECG waveform, and C type is a signal which has obvious baseline drift and is clinically unavailable with large noise.