Intra and inter-patient arrhythmia classification using feature fusion with novel feature set based on fractional-order and fibonacci series

https://doi.org/10.1016/j.bspc.2021.103365Get rights and content

Abstract

Cardiac ailments are increasing at an alarming rate globally due to the sedentary lifestyle and increased desk-bound activities. For decades, the ECG signal is used in aiding the analysis of human heart. Arrhythmia is a disorder that alters the normal cardiac cycle of ECG signal. The automatic classification of arrhythmia is a highly desirable and tough task. Mainly, the traditional methods of arrhythmia classification, are evaluated on the intra-patient criterion that may not befit the inter-patient criterion. A complete classification method has been proposed in the paper which performs well in intra-patient, inter-patient criterion and it is effective for minority class of MIT-BIH arrhythmia database (MIT-BIH-AD). In the proposed method, after preliminary processing based on Riesz fractional-order digital differentiator and R-peak detection, various features are extracted from ECG beat. The feature set proposed is a fusion of time-domain, time–frequency domain features and the new and novel features based on the Fibonacci series and coefficients of fractional-order Riesz based derivative signals which are proposed in this work. The results are evaluated on the MIT-BIH-AD and the results for intra-patients’ criterion have achieved an overall accuracy, sensitivity and positive predictivity values of 99.85%, 99.16%, 99.93% respectively, and for the inter-patient scheme 92.5%, 89.89% and 95.54% an average value of accuracy, sensitivity and positive predictivity of six ECG classes. The obtained results in both criterions have outperformed other methods in the literature and the proposed work has also attained better results for minority class of MIT-BIH-AD in both the scheme.

Introduction

Cardiac arrhythmia is the typical cardiac disorder that affects the normal functioning of the heart which may cause temporary ailments or sudden death. The timely and accurate prediction of it can save a lot of life. So, long term ECG signal is required and this job automation is tremendously preferable [1], [2]. The automation method for classification comprises of four parts (i) ECG signal processing, (ii) beat segmentation, (iii) extraction of the feature, and (iv) classification. The first two parts are extensively investigated by the research scholars. There is still scope for improvement in the last two parts.

Extraction of the feature is an essential step in the image and signal classification process. The features are procured from the raw ECG signal and transformed ECG signal. Features can be morphological, statistical, or both. The most frequently used ECG feature is the RR interval that is described as the time interval between two adjacent R-peaks of the ECG record [3], [4]. Other features are based on various heartbeat intervals and segments and ECG morphology [5]. Various algorithms are proposed as in [5], [6], [7] to determine the fiducial points in ECG and to determine the values of various heartbeat intervals as healthy heartbeat intervals fall in a certain range.

The other statistical approach to extract feature is independent component analysis (ICA) [8]. ECG signal was decomposed into a weighted sum of basic components (mutually statistical independent) using ICA. A new feature set of ECG beat was developed and used in [9] which includes morphological features and ICA extracted features.

Another significant technique to extract features is from the transformed domain. The most commonly used approaches are discrete wavelet transform (DWT), S-Transform (ST), and Empirical mode decomposition (EMD). Wavelet was commonly used as it can extract discriminative features for ECG classification. The dimensionality reduction algorithms principal component analysis (PCA), linear discriminant analysis (LDA) and independent component analysis (ICA) was employed on fourth level detail and approximation coefficients of DWT independently and six features from both sub-bands were extracted and fed to classifiers [10]. ST based features are extracted and a method combining of ST and DWT based features is proposed in [11]. ECG beat is broken down into intrinsic mode functions (IMF) using two approaches: EMD and ensemble empirical mode decomposition (EEMD) [12]. Four features were extracted from the obtained IMFs from both the approaches independently and fed to the classifiers for ECG classification. This technique showed its ability in differentiating different heartbeats even under noisy conditions.

As for ECG beat classification, any multi-class classifier is used. The frequently employed models are support vector machine (SVM) [13], [14], artificial neural network (ANN) [15], K-nearest neighbor (KNN) and decision tree (DT) [4].

IMFs based features obtained in [12] were fed to a sequential minimal optimization-support vector machine (SMO-SVM). A method for an ensemble of SVMs is proposed in [14]. This approach has performed better than the single SVM model when the same feature set was used. In [2], the author fine-tunes the SVM classifier and performed feature selection using a particle swarm optimization (PSO) algorithm. The other classifiers employed are hidden markov model (HMM) [16], multilayer perceptron neural network (MLPNN) [11] and random forest [17], and so on to build classification models.

Extensive research on the classification of ECG arrhythmia exists, but further exploration of this field is still essential. (i) the existing research although has achieved good accuracies but they mainly focus on intra-patient criterion or beat-based scheme (ECG beats of the same subject is used in training and testing process) whereas, in a practical scenario, training set would be from different subjects and the testing set would be from entirely different subjects (inter-patient criterion or record-based scheme) [4]. (ii) the prevailing methods have poor performance for imbalanced datasets, as they achieved poor results in terms of sensitivity and positive predictivity for small size class in the dataset. For example, for class atrial premature contraction (A) in the MIT-BIH [18] which comprises 2.37% of the total dataset (mentioned in section 2).

However, addressing the above mentioned issues is a major motive of this study. Our work focuses on extracting informative and non-redundant features for better classification results. As fractional-order calculus (FOC) secures the signal information more successfully than the integer-order calculus (IOC) [19], [20], [21] and our previous work has also revealed this in the context of ECG signal processing [19]. This paper focuses on extracting ECG features that reflect more details and intrinsic features about ECG beat. The proposed approach employs new and novel features for ECG beat extracted based on the Fibonacci series, which was used earlier for the classification of acoustic signals [22] and Riesz based fractional order filtering [19]. It is the first time, as far as we know, that the Fibonacci series and fractional-order features are used for arrhythmia classification. The goal of this work is, therefore, to handle this pivotal and tough task, and authenticate the obtained results with the established work.

Section snippets

Material

Many ECG databases are available to validate the newly developed approaches. The use of authenticated databases is highly encouraged as the results obtained are equivalent to real-world signals. The MIT-BIH arrhythmia database (MIT-BIH-AD) is one of the commonly used and the most suggested database for evaluating automatic arrhythmia classification. This database has 48 records of ECG signal each 30 min duration. ECG beats in this database has changing QRS morphology, inverted R-peaks, inverted

Proposed methodology

The algorithm proposed for ECG arrhythmia classification is shown in Fig. 1. The raw ECG is procured from the MIT-BIH-AD. There is a common framework for every classification system. It comprises pre-processing, extraction of feature, and data classification. Pre-processing of the ECG signal is a very important step as it prepares the waveform to facilitate and improve the extraction of features that best represent the ECG beat. In a real-world scenario, noises and artifacts such as, baseline

Conclusions

This work proposes a new and novel technique for classifying ECG arrhythmia. The suggested work amalgamated the Riesz based fractional-order derivative signal coefficient, Shannon-energy of the Fibonacci signal obtained from the beat with other time-domain and time–frequency features. The proposed method proves to be superior than the other methods in literature as it has achieved promising performances in challenges faced in the literature (i) it performed well for the minority class i.e. A of

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

The work was received support from Science and Engineering Research Board (SERB) (No. SB/S3/EECE/0149/2016), Department of Science and Technology (DST), Government of India, India.

References (34)

Cited by (5)

  • Detection and Classification of Arrhythmia Using Hybrid Deep Learning Model

    2023, 2023 International Conference on Next Generation Electronics, NEleX 2023
  • CLSTM-Transformer: Inter-patient ECG classification for ventricular arrhythmia

    2023, Proceedings of SPIE - The International Society for Optical Engineering
View full text