RR Interval-based Atrial Fibrillation Detection using Traditional and Ensemble Machine Learning Algorithms

Abstract

Atrial fibrillation (AF) is a life threatening disease and can cause stroke, heart failure, and sometimes death. To reduce the rate of mortality and morbidity due to increased prevalence of AF, early detection of the same becomes a prior concern. Traditional machine learning (TML) algorithms and ensemble machine learning (EML) algorithms are proposed to detect AF in this article. The performances of both these methods are compared in this study. Methodology involves computation of RR interval features extracted from electrocardiogram and its classification into: normal, AF, and other rhythms. TML techniques such as Classification and Regression Tree, K Nearest Neighbor, C4.5, Iterative Dichotomiser 3, Support Vector Machine and EML classifier such as Random Forest (RF), and Rotation Forest are used for classification. The proposed method is evaluated using PhysioNet challenge 2017. During the tenfold cross validation, it is observed that RF classifier provided good classification accuracy of 99.10% with area under the curve of 0.998. Apart from contributing a new methodology, the proposed study also experimentally proves higher performance with ensemble learning method, RF. The methodology has many applications in health care management systems including defibrillators, cardiac pacemakers, etc.

Introduction

Atrial fibrillation (AF) is one of the most commonly observed types of cardiovascular disease which is clinically important. Worldwide, it occurs with an incidence and prevalence rate of 0.4% and 2.75% per year, respectively.[1,2] According to the recent report of American Heart Association, more than 33 million individuals are affected by AF worldwide.[3,4] Every year 5 million new cases of AF are being reported globally.[5] The statistics of AF indicate that the incidence of AF is progressively increasing with years. Furthermore, it is noted that the incidence of AF is higher in the elderly population. The prominent causes of AF include obesity, physical inactivity, increased intake of fatty food, smoking habit, consumption of alcohol, and family inheritance. Clinically, AF can easily be diagnosed by the electrocardiogram (ECG) pattern. Usual findings of an individual's ECG with AF include irregularity in the R peak to R peak time interval followed by the absence of P wave.

Many authors have proposed a number of computational techniques and algorithms applying Machine Learning (ML) and Deep Learning (DL) approaches in diagnosing AF using digital ECG signal. Graphical abstract of the proposed approach is shown in Figure 1. Table 1 provides the literature review of some prominent work related to ECG classification.

Figure 1.

Graphical abstract of the proposed approach

Table 1.

Literature review

Literature	Techniques	Datasize	Validation	Classifier	Overall accuracy
Luo et al.[6]	Poincare plot	229 ECG signal from different database	1 fold	Threshold	Sensitivity of 97.91% and specificity of 99.14%
Radhakrishnan et al.[7]	Chirplet transform	34206 ECG signals from 3 database	10 fold	CNN-BiLSTM	99.18%
Shi et al.[8]	Canonical correlation	8528 ECG signals	1 fold	CNN	91.7%
Nurmaini et al.[9]	DWT	5504 Training signals 824 Testing signals 612 Validation signals	10 fold	1D-CNN	99.98%
Lown et al.[10]	Lorenz plot	19749 ECG signals from different database 415 Validation signals	1fold	SVM	Sensitivity 100% specificity 97.6%
Liang et al.[11]	Deep features	7877 ECG signals	10 fold and random oversampling	CNN - BiLSTM	85%
Ghosh et al.[12]	Fractional norm feature	190 ECG signals 30 testing signals	10 fold	Hierarchical extreme learning	99.40%
Wu et al.[13]	Deep features	5546 ECG signals	10 fold	RF	F1 score of 96%
Wang et al.[14]	Wavelet packet transform	141556 ECG segments	10 fold	Artificial neural network	98.8%
Jin et al.[15]	Residual block	92789 ECG signals from different data base	1 fold	CNN	98.84%
Yue and Jinjing[16]	Empirical mode decomposition	8528 ECG signals	5 fold	Gradient boosting classifier	86%
Horoba et al.[17]	RR interval	25 ECG signals of 10 hour	10 fold	SVM	94%
Kong et al.[18]	RR interval	1960 ECG signals	10 fold	SVM	98.16%
Guo et al.[19]	Spectrogram	8528ECG signals	1 fold	CNN	78%
Aligholipour et al.[20]	Morphological features	100 ECG signals	4 fold	Neural network	93.05%
Firoozabadi et al.[21]	Self organizing map	8809 ECG signals from different database	5 fold	Decision tree	F1 score of 96%
Maji et al.[22]	Empirical mode decomposition	165 ECG signals	5 fold		Sensitivity of 96.14 % and specificity of 93.51%.

ECG – Electrocardiogram; DWT – Discrete wavelet transform; CNN – Convolutional neural network; SVM – Support vector machine; RF – Random forest

Luo et al.[6] proposed polar coordinate transformation method on Poincare plot and recorded specificity of 99.14%. Radhakrishnan et al.[7] proposed chirplet transform method using deep convolutional Bidirectional Long Short-Term Memory (BiLSTM) network and achieved accuracy of 99.18%. Shi et al.[8] suggested convolutional neural network (CNN)-based method to detect AF by extracting deep features and RR interval features and recorded accuracy of 91.7%. Nurmaini et al.[9] proposed discrete wavelet transform (DWT) along with 13 layers of 1 dimensional CNN to discriminate ECG signal and recoded F1 score of 99.98%. Lown et al.[10] developed AF detection algorithm by using Lorenz plot of RR intervals. They have also used wavelet transformation to compress the image of ECG signal and is given as input to Support Vector Machine (SVM) classifier and recorded sensitivity and specificity of 100% and 97.6%. Liang et al.[11] suggested the combination of CNN and BiLSTM classifier extracting deep features to classify heartbeat events and observed accuracy of 85%. Ghosh et al.[12] suggested hierarchical extreme ML approach for the diagnosis of AF using multirate cosine bank filter combined with fractional norm feature and observed accuracy of 99.40%. In a work by Shankar et al.,[23] various dimensionality reduction techniques, features used and different ML methods are systematically reviewed. Kleyko et al.[24] conducted a study on computational complexity of automatic detection of AF and classification using different databases. Wu et al.[13] proposed deep features-based approach to diagnose AF using Random Forest (RF) classifier and recorded F1 score of 96%. Wang et al.[14] suggested wavelet packet transform method for efficient feature extraction and detection of AF. The author used artificial neural network for the classification purpose and recorded accuracy of 98.8%. Jin et al.[15] proposed CNN method combined with residual block method to detect AF and recorded accuracy of 98.84%. Yue et al.[16] designed ensemble empirical mode decomposition filter and extreme gradient boosting classifier to identify AF signal and observed accuracy of 86%. Horobo et al.[17] proposed an approach based on RR interval features using Lagrangian SVM classifier to detect AF and reported accuracy of 94%. Kong et al.[18] proposed RR interval-based method for the diagnosis of AF using relevance vector machine algorithm and observed accuracy of 98.16%. Guo et al.[19] suggested a algorithm that incorporates spectrogram and CNN for the diagnosis of AF with accuracy of 78%. Aligholipour et al.[20] developed neural network-based algorithm using nonlinear method-based features to identify AF and recorded accuracy of 93.05%. Hagiwara et al.[25] reviewed different computer-aided diagnosis methods consisting of ML algorithms and DL algorithms to detect AF developed by various researchers. Firoozabadi et al.[21] suggested self-organizing map technique for atrial and ventricular activity features to diagnose AF using decision tree to record F1 score of 96%. Maji et al.[22] proposed a methodology to detect AF using the empirical mode decomposition method with sensitivity of 96% and specificity of 93%.

With close observation of existing literature on AF detection, made by authors using Traditional ML (TML) methods or DL methods, yet an exhaustive experimentation and corresponding broader comparative analysis with respect to effective performance does not exist. Different authors have designed experiments with different dataset, varied data size, different split of training, and test data during cross validation, but the comparison of F1score of these experiments is not logical.

The contribution of the present study is the usage of ensemble learning method and verification of experimental proof that ensemble method performs better with high accuracy. The present study involves the extraction of RR series features from ECG signals from the Physionet Challenge 2017 database and training of TML classifiers such as classification and regression tree (CART), K-Nearest Neighbor (KNN), C4.5, Iterative Dichotomiser (ID3), and SVM and ensemble ML (EML) classifiers such as RF, and Rotation Forest and reporting of the obtained performance in the classification of ECG signal into normal, AF, and other rhythms. The proposed approach is illustrated in section II, the results obtained and their significance are discussed in section III, and the paper concludes with section IV.

Methodology

The proposed approach is shown in Figure 2. It involves the removal of baseline wander and other noise components by using DWT, R peak detection using Pan Tompkins algorithm, RR series feature extraction and features selection and finally classification into normal, AF, and other rhythms. Each of the used method in the proposed approach is explained in this section.

Figure 2.

Proposed approach

Data set used

In this study, open source dataset, Physionet Challenge 2017, is used. The dataset used in this study comprises of 5154 signals with normal rhythm (N), 771 signals with AF rhythm, and 2557 signals with other rhythms. Each of the ECG signals has varied length between 9 s and slightly more than 1 min.[26]

Preprocessing

ECG signals include distinct noises such as baseline wander, muscle and movement artifacts, interference of power line, electrode pop noise, etc., DWT is used to remove the noises present in the ECG signal. Daubechies6 (db6) is used as basis function in the time frequency decomposition of ECG signal using DWT.[27] The ECG signal sampled at 300 Hz is decomposed into eight levels. The frequency band from 75 Hz to 150 Hz does not include the required components of the ECG signal. This sub-band represents the I^st level detail. Hence, this sub-band is not considered during the reconstruction step. The frequency component between 0 and 0.5 Hz represents baseline wander and is the eighth level approximation sub-band which is also not required. The necessary bands include IInd, IIIrd, IVth, Vth, VIth, and VIIth level's detail sub-bands. Hence, first level detail coefficients and VIIIth level approximation coefficients are replaced with zeros during the reconstruction process of inverse wavelet transformation to obtain the denoised ECG signal. The denoised ECG signal is further subjected to QRS complex detection with the help of Pan Tompkins algorithm[27,28] to detect RR intervals.

Feature extraction

Various features are extracted which are enumerated[25,29,30,31,32] in Table 2.

Table 2.

Time domain features extracted from electrocardiogram signal

Feature	Computational equation	F	P	Rank
RRµ	Mean of RR intervals RRµ= Where RRi is the RR interval at ith instant and N is the length of RR interval	8.02	0.0003	9
CVRR	Coefficient variance of RR intervals CVRR=	1.29	0.2753	15
SDRR	Standard deviation of RR intervals SDRR=	3.79	0.227	13
SDSD	Standard deviation of RR interval differences SDSD=STD (RRi-RRi+1) Where STD is standard deviation	5.38	0.0046	10
RMSSD	RMS of successive differences RMSSD=	2.93	0.0533	14
RR50	Number of pairs of adjacent RR intervals differing by 50 ms RR50=(RRi-RRi+1)>0.05	3.92	0.0198	11
pRR50	Proportion successive RR interval >50 ms pRR50=100*RR50/Length of RR intervals	9.12	0.0001	7
Triang8	It is the total number of RR intervals divided by the height of histogram in 8 m bins	13.77	1.07e-04	5
TINN8	Multilinear function is q is defined such that q (t)=0 for t <A and t >B and q (X)=Y TINN8=B-A where 8 stands for 8 ms bins	17.72	2.09e-08	3
pRR20	Proportion successive RR interval >20 ms pRR20=100*(RRi-RRi+1) >0.02/(Length of RR intervals)	21.21	6.47e-10	1
pRR30	Proportion successive RR interval>30 ms pRR30=100*(RRi-RRi+1) >0.03/(Length of RR intervals)	15.08	2.88e-07	4
pRR6.25	Proportion successive difference >1/16 ms pRR6.25=100*(RRi-RRi+1) >0.00625/(Length of RR intervals)	17.95	1.66e-08	2
RSA 5RR	Difference betweenmean of 5 largest and5 smallest RR intervals	9.9	5.09e-05	6
SampEn	Sample Entropy SampEn (m, r, N)= where D=Number of guided vector pairs having distance function d[Xm+1(i), Xm+1(j)]<r E=Number of guided vector pairs having distance function d[Xm (i), Xm (j)]<r	2.8	0.00061	12
ApEn	Approximate entropy ApEn= where Brm (i)=[number of x (j) such that d (x(i)), x (j)<r/(N-m+1)	8.16	0.0003	8

Feature selection

The selection of appropriate features plays an important role in enhancing the accuracy of classification and also reducing the computational time. In this study, we have used the Analysis of Variance (ANOVA) test to select features. The value of F and p obtained from ANOVA test is enumerated in Table 2. The relative importance of each feature is expressed in terms of rank.

Classification

In order to discriminate the ECG signal into 3 rhythms viz: normal, AF and other rhythms, TML classifiers such as CART, KNN, C4.5, ID3, and SVM and EML classifiers such as RF, and Rotation Forest are used.

Classification and regression tree

CART is a nonparametric supervised learning method used for classification. CART makes use of If and Else rules to arrive at a particular condition. It is treelike structure and has three nodes, namely root node, child node, and leaf node. Information gain is calculated at each node. This information gain is used for the further division of the tree. This division continues till leaf node, where the information gain becomes zero.[27]

K-Nearest Neighbor

K-NN is a nonparametric and supervised ML algorithm. It works on the concept of similarity measures. It associates the vector of given feature with the category that is more similar to the available categories. The similarity can be calculated using distance measurements. In the present study, we have used Euclidean distance which is given by.[28]

Support vector machine

In the present study, the SVM classifier is used to classify heart rate variability derived features into normal, AF, and other rhythms. SVM classifier works on the principle of constructing a hyperplane which linearly separates two classes.[33,34] A hyperplane can be constructed using

In our study, we have used Radial Basis Kernel function which is given by

Where (x_i, y_i) represent training set, x_i∈ Rd, i = 1,2,….l and y_i∈ (−1,1)

Rotation forest

Rotation forest is a decision tree-based ensemble classifier. The rotation forest increases the accuracy of the individual tree, yet maintains the diversity within the ensemble. The RF algorithm is sensitive to the rotation of feature axes.[35] The final average performance of rotation forest is the average performance of all individual trees. The rotation forest can achieve better performance with less number of trees. It is more immune to noise present in the data. In the present study, the rotation forest is used to classify HRV derived features into normal, AF, and other rhythms. As the rotation forest is robust to noise and as it provides a simple tree structure with minimum number of parameters, it is expected to provide relatively higher performance.[36]

C4.5

C4.5 is a decision tree-based algorithm and is an extension of Quinlan's ID3 algorithm used for classification of data. C4.5 constructs decision tree from the set of training data using the concept of information entropy. C4.5 selects the attribute of the data at each tree node that most efficiently splits its sample set into subsets enriched in one or the other class.[37] The splitting criteria are based on the difference in entropy of data. To make the decision, the attribute with the highest difference in entropy is selected. The C4.5 algorithm then splits by selected attribute to produce remaining on the sub-list.

Iterative dichotomiser 3

ID3 algorithm is used in ML and natural language processing. It is used to generate the decision tree from the given dataset. ID3 algorithm begins with the original set S as the root node. At every iteration of the algorithm, it iterates through every unused attribute of the data set S.[38] After iteration, it calculates entropy of the attribute. Algorithm then selects the attribute which has smallest entropy value. The set S is then split by selected attribute to produce sublist of the data.

Results and Discussion

In this study, a three class pattern classification problem of detection of normal, AF, and other rhythms using the ECG derived RR interval features is proposed. The methodology is implemented in MATLAB using Physionet Challenge 2017 dataset. Baseline wander and high frequency noise present in the ECG are removed using DWT. Pan-Tompkins algorithm is used to detect QRS complex from denoised ECG. Figure 3 (a) shows raw ECG consisting of baseline wander. Figure 3 (b) shows denoised ECG using DWT where the baseline wander is removed. Figure 3 (c) provides the visualization of detected QRS complex which is marked in red asterisk from the denoised ECG. Various features, extracted from the RR interval time series, which are listed in Table 2. It can be seen that each feature has a different F and p indicating different discrimination potential and statistical significance, respectively, during ANOVA test. In total, the extracted 15 features are subjected to classification using TML algorithms as well as EML algorithms.

Figure 3.

(a): Raw ECG. (b) Denoised ECG using DWT. (c): QRS complex detection. ECG: Electrocardiogram, DWT: Discrete wavelet transform

The average 10-fold cross validation results and area under the curve (AUC) for CART, KNN, RF, Rotation Forest, C4.5, ID3, and SVM classifier are shown in Table 3. It can be observed that RF classifier achieved an overall accuracy of 99.10%.

Table 3.

Average classification performance of classification and regression tree, K-nearest neighbor, random forest, Rotation Forest, C4.5, iterative dichotomiser, and support vector machine classifier

Classifier	NCSA±SD	AFCSA±SD	OCSA±SD	OA±SD	AUC±SD
CART	98.64±0.44	97.86±0.71	97.94±1.24	98.34±0.43	0.972±0.0066
KNN	98.24±0.89	92.67±3.66	93.30±2.99	96.16±1.75	0.968±0.0060
RF	99.65±0.18	99.50±0.18	97.98±0.47	99.10±0.25	0.998±0.0006
Rotation Forest	99.27±0.24	97.81±0.78	97.29±0.69	98.50±0.38	0.976±0.0059
C4.5	99.19±0.18	98.35±0.60	97.20±0.28	98.47±0.23	0.979±0.0037
ID3	98.91±0.09	98.31±1.03	96.98±1.28	98.24±0.49	0.989±0.0042
SVM	96.18±0.32	97.26±0.38	95.56±0.64	96.16±0.54	0.956±0.0032

NCSA – Normal class specific accuracy; AFCSA – AF class specific accuracy; OCSA – Other rhythm class specific accuracy; OA – Overall accuracy; AUC – Area under the curve; CART – Classification and regression tree; KNN – K-nearest neighbor; RF – Random forest; ID3 – Iterative dichotomiser; SVM – Support vector machine; SD – Standard deviation

From Table 3, it could be concluded that the RF classifier provides higher performance metrics (NCSA of 98.64%, AFCSA of 99.50%, OCSA of 97.98%, and OA of 99.10%) in comparison with other classifiers. During the training and the testing of the classifier, the 10 fold cross validation is used. It can be seen that RF classifier provides the highest overall accuracy with a minimum variation in respect to different folds. From the classified signals receiver operating characteristic curve is plotted to calculate AUC. It is observed that AUC of RF classifier is 0.998 which is more than the other classifiers.

The scatter plot is presented in Figure 4 for the first 2 features of 100 signals. The various distinct regions resulted due to classifier training can be seen in the scatter plot.

Figure 4.

Scatter plot of first two features with class boundary during RF classification. RF: Random forest

In the present study, the RF classifier provided higher performance in comparison with other EML and TML classifiers. It is interesting to compare the performance obtained in the present study with the existing state-of-the-art methods available in the literature. Table 4 highlights some of the work carried out by authors on detecting AF using Physionet challenge 2017 dataset. Rao et al.[27] proposed effective noise removal method along with dimensionality reduction techniques using CNN to detect AF and recorded accuracy of 91.71%. Quang et al.[33] suggested a method based on statistical features of segments derived from CNN and recorded F1 score of 84.19%. Plesinger et al.[39] proposed a method to classify ECG signal using PQRS morphology features using Bagged tree ensemble and shallow NN and achieved F1 score of 91%. Liu et al.[40] suggested a method using SVM classifier for discrimination of ECG signal using P wave, RR interval, spectrum, entropy features, and recorded average F1 score of 90.82%. Chen et al.[41] proposed morphological and heart rate variable features method to detect AF. Piecewise linear splines method was used for the selection of features and recorded F1 score of 81%. Mei et al.[42] proposed heart rate variability and frequency features for the diagnosis of AF using SVM classifier and achieved accuracy of 96%. Shao et al.[43] proposed delta RR interval, morphology, similarity index feature to detect AF using AdaBoost classifier and recorded F1 score of 82%. Sanchez et al.[44] suggested gramian angular summation field method to detect AF using CNN and achieved accuracy of 97.6%. Najmeh et al.[45] suggested neural architecture search method to diagnose AF using CNN classifier and observed F1 score of 84.15%. Rao et al.[46] proposed power spectrum-based method for the classification of ECG signal using CNN and reported 94.67% of accuracy.

Table 4.

Overview of studies on classification of electrocardiogram signal using physionet challenge 2017 database

Literature	Techniques	Classifier	Overall accuracy
Shrikanth et al.[27]	PCA and ICA	CNN	91.71%
Nguyen et al.[33]	Stacking method	SVM and CNN	F1 score of 84.19%
Plesinger et al.[39]	PQRS morphology features	Bagged tree ensemble and shallow neural network	F1 score of 91%
Liu et al.[40]	P wave, RR interval, spectrum, entropy features	SVM	F1 score of 90.82%
Chen et al.[41]	Morphological and heart rate variability features	XGBoost	F1 score of 81%
Mei et al.[42]	Heart rate variability and frequency features	SVM	96%
Shao et al.[43]	Delta RR interval, morphology, similarity index features	AdaBoost classifier	F1 score of 81%
Sánchez et al.[44]	Gramian angular summation fields	CNN	97.6%
Fayyazifar et al.[45]	Neural architecture search	CNN	F1 score of 84.15%
Rao et al.[46]	Power spectrum	CNN	94.67%
Current Study	RR series feature	RF	99.10%

CNN – Convolutional neural network; SVM – Support vector machine; RF – Random forest; PCA – Principal component analysis; ICA – Independent component analysis

The current study focuses on extracting RR interval based features and their classification using TML and EML classifiers. The 15 features are extracted from RR interval series are subjected to classification using CART, KNN, RF, Rotation Forest, C4.5, ID3, and SVM classifier. During the study, it is found that RF performed better compared to other classifiers achieving an overall accuracy of 99.10%.

It is noted that there are many studies in the literature which classify the ECG into different classes (as seen in Table 4). Innovation in this proposed study is the usage of ensemble learning method and its superiority in terms of high classification performance in accurately categorizing the ECG data into normal, AF and other rhythms. Apart from the methodological contribution, the method also provides improved performance, which is the key contribution of this study. Referring Table 4, all other authors in previous studies obtained relatively lower performance compared to the proposed study.

The proposed study has the application in monitoring of patients in hospitals, critical care units, Holter monitoring of patients during their normal activity, pacemakers, and defibrillators. The developed system can be used in mass screening of population, remote patient monitoring, and in many other applications. Automated computer diagnosis of AF can bring a change in the conventional practice by immensely helping the physicians in accurately diagnosing and mobilizing patients.

Conclusion

In this work, we propose RR interval features-based methodology to detect AF. Physionet challenge 2017 database is used in this study. DWT method is used for the denoising of ECG signal. TML algorithms and EML algorithms are used to evaluate the performance. It is observed that RF classifier gave a better discrimination with NCSA of 99.65%, AFCSA of 99.50%, OCSA of 97.98%, and OA of 99.10%.

There is an increase in the number of patients with heart arrhythmia, worldwide. Manually diagnosing each arrhythmia and its classification into different categories is extremely tedious and time consuming for any physician. An accurate method of classification of ECG signal into various rhythms is very important. Each abnormality requires different method of treatment and any misclassification may cause serious complications. In this direction, obtaining high performance is highly desirable. The proposed methodology provided an improved accuracy on the benchmark dataset. The developed methodology can be used in the health care management system to screen ECG signal into various rhythms.

Financial support and sponsorship

None.

Conflicts of interest

There are no conflicts of interest.