Detection of Cardiac Arrhythmia Using Multi-Perspective Convolutional Neutral Network for ECG Heartbeat Classification

Maintenance of health is the most crucial part of human life as it states the quality of life of a person [1]. Out of all the organs in the body, the most important organs that plays key role for human life and body functionality is ones heart. The blood throughout the body is pumped through heart. The vital functionality of heart is to pump blood throughout the organs of the body, so that healthy maintenance of heart is very important. When there is abnormality in functioning of heart, the worst experience a person can have is death. According to 2016 census of the World Health Organization (WHO), the number of people dies because of heart disease 17.9 million i.e.,  more or equal to 31% of the world population died due to heart diseases. It is also identified that under 70 years of age around 17 million people were dead. As per the WHO sources The rhythm of heart beat can be deviated from its electrical activity [2] which is considered as Cardiac Arrhythmia. There are various kinds of Arrhythmia which are harmless. But, when Arrhythmia is caused due to the weakness of damaged heart, it may lead to serious consequences and fatal symptoms [3]. Usually Arrhythmias is classified into two kinds the first one is Brady cardiac which means slow heart beat and the other is tachycardia which means fast heartbeat. When a person has complaints relating to heart functioning, as a part of initial assessment electrocardiogram (ECG) is performed. The ECG device represents the entire process of polarization and re polarization in the form of X-ray sheets.

The structure and functioning of heart can be studied with the help of Electrocardiography (ECG) which is an efficient technique used due to ease of access, non-invasiveness and less expensive [4]. During each heart beat it presents the depolarization and re polarization of muscles of heart is shown in the form of electrophysiological pattern. Due to various factors like baseline wander, external noise, physical variations among individuals [5] the heart beat classification automatically using ECG has become a challenging task. In case of a healthy person also there may be difference in rhythm and morphology of heartbeats during various circumstances [6]. Few elements that are vigorous against such situations are used in classification of heartbeat like characteristic points relating to morphological features, abstract features and transformation features generated by seizure interpretation.

In the field of biomedical signal processing and digital image processing [7] shows high demand for the deep learning algorithms which involves structures such as neural networks. A Machine Learning (ML) based technique called Deep Learning (DL) has the capacity of automatic feature extraction. A layer of favorable features are determined from the input data using deep learning without involving any feature extraction method [8]. Analysis of physiological information, image categorization and speech identification are the traditional practices. Compared with these conventional methods effective results are exhibited by using Deep Learning techniques. Convolutional Neural Network (CNN) [9] is a branch or an application of deep learning. Two shape dimensional data [10] can be processed by using CNN technique.

Different data classification techniques such as wavelet transform, Artificial Neural Networks (ANN), Support Vector Machine (SVM) and hidden Markov models were developed previously for automatic detection of ECG data classification approach. Signal pre-processing and Feature extraction are the important features of these techniques. The medical expert is required for feature extraction by using hand-crafted methods [11, 12].

As these techniques are performed manually they consume lot of time, costly and there may be chance of losing data during the phase of extracting features. Therefore these features faced lot of challenges in identification of arrhythmia symptoms may exhibit different morphological signal during different circumstances [13]. Due to these elements, the performance classification was not achieved when applying new ECG data.

The various sections in this paper are well thought-out as follows: Section II refers to the literature survey. Section III refers to the detection method of Cardiac Arrhythmia using Multi-Perspective Convolutional Neutral Network for ECG Heartbeat Classification. Section IV presents the evaluation of experiment and its results. Section V explains about conclusion and works to be performed and developed in future.

Overall framework of detection of Cardiac Arrhythmia using Multi-Perspective Convolutional Neutral Network for ECG Heartbeat Classification is represented in below Figure 1.

1.png

Figure 1. Block diagram of ECG heartbeat classification

The MIT-BIH Arrhythmia based Database is utilized in data processing of this paper as represented in the above Figure 1. The ECG signals are classified and data augmentation is performed as presented in the above figure. This dataset was recorded from 48 individual subjects and it consists of two channels ECG signals of total 48 records which is recorded around the duration of 30 minutes. For the performance investigation of 1D convolutional neural network model, a total of 109446 beats at 125 Hz sampling frequency are collected from 44 records for evaluation of testing and training patterns in this research paper. When the signal features are not adequate for sound processing they are eliminated from the evaluation process.

Basing on the suggestion from AAMI, ECG heartbeat can be classified into 5 heartbeat types they are as follows:

5. N–Normal Beat

5. SVP–Supra Ventricular Premature beat

5. PVC–Premature Ventricular Contraction Beat

5. FVN–Fusion of Ventricular and Normal Beat

5. FPN–Fusion of Paced and Normal Beat

To produce the ECG beats from a specified ECG waveform, there are dissimilar forms of preprocessing required to improve the ECG signal efficiency enhancement. As ECG signals involves various forms of destructive noises, few steps are suggested and adopted for ECG signal database processing in three phases, the first one is de-noising the ECG dataset and then QRS peak detection is performed which is a remarkable waveform in ECG diagnosis and later, in the final step heartbeat signals are segmented in different clusters for each single beat of an individual is performed.

The entire dataset must be augmented to the same level for training the model effectively. This can be performed by selecting the smallest heartbeat sample classes. The data is down sampled due to the imbalance of the dataset that results in misclassification. The raw data of individual heartbeats is represented with fixed dimensions of 188 samples.

Later, Symbolic Baseline Corrected approximation (SBCX) is performed in the next step. In this phase, SBCX explains about the consideration specially by using symbolization method widely for a particular time series known as Symbolic Aggregate approximation (SAX). Symbolization and dimension reduction are the two steps involved in SAX.

The input heartbeat signal can be examined for a baseline approximation by the SBCX. By using a single heartbeat signal, heuristic finding of ECG signal baseline is a challenging task. Basing on the nature of the stately biomedical signal, an effective approach is designed. The input signal of heartbeat is taken as x=x₁, ..., x_i, ..., x_n when normalized with the global maximum value (v_max). Later, the signal amplitude of the global minimum value (v_min) of training data is given as follows:

$x_i^{\mid}=\frac{x_i-v_{\min }}{v_{\max }-v_{\min }}$      (1)

Baseline Correction Normalization: A baseline corrected normalization method is introduced using mode $\left(x^{\mid}\right)$ as the approximate baseline.

$\sigma_{m o d e}=\sqrt{\frac{\sum_i\left(x_i^{\mid}-\operatorname{mode}\left(x^{\mid}\right)\right)^2}{n}}$       (2)

$z_i=\frac{x_i^{\mid}-\operatorname{mode}\left(x^{\mid}\right)}{\sigma_{\text {mode }}}$       (3)

After baseline correction, the signal values are divided by σ mode in order divide the heartbeat signals into various segments having similar “scale of amplitude”. The normalized signal z=z₁….z_i….z_n step is performed.

After completion of the normalization process, a breakpoint list α and a corresponding vocabulary list A are used for symbolization as represented in the following Eq. (4).

$z_i^{\mid}=a_k$, if $\alpha_{k-1} \leq z_i<\alpha_k$           (4)

The final SBCX representation is $z^{\mid}=z_1^{\mid}, \ldots . z_i^{\mid} \ldots . z_n^{\mid}$ where $a_k \in A$ A is the mapped signal symbol. SBXC is a sequence of symbol.

ECG signal is transformed into symbolic representation of SBCX by using the SBCX method. The embedding vectors are present for all symbols in lookup table which is a trainable matrix. The symbol sequence of the signal is further transformed into an embedding matrix represented as $S_{\text {signal }} \in R^{d \times n}$.

$S_{\text {signal }}=\left[\begin{array}{ccccc}\mid & & \mid & & \mid \\ S_1 & \cdots & S_i & \cdots & S_n \\ \mid & & \mid & & \mid\end{array}\right]$          (5)

where, $S_i \in R^d$ is the vector that is mapped to $z_i^{\mid}$ and the embedding dimension is represented as ' $\mathrm{d}$ '.

The SBCX parameters are nontrainable and grid searching is performed for determining them. The various signal vector values in A are derived from the trainable weights from the entire model. During training stage, they are randomly initialized and the MPCNN weights are optimized simultaneously by minimizing the loss function. The heartbeat feature extraction in embedding matrix is performed by Multi Perspective Convolutional Neural Network (MPCNN) and the kind of input heartbeat is predicted. Such kind of layer combination is observed as an effective method in various applications.

2.png

Figure 2. Architecture of MPCNN

The Figure 2 explains the detailed architecture of Described MPCNN, which consists of variety of channels that consists of various stack modules. The signal amplitude consists of information which is directly exploited by the shallow channels such as M11, M21, and M41. In the same way high level features are extracted by the deeper channels such as M51-M52 and M31-M32 in relation with the global shape information and morphological structure. The type of heartbeat can be predicted by concatenating the multi perspective representations and feeding them into a softmax layer. It is to be noted that deeper channels have to be tried in beginning of the experiment. From the experiment results it can be observed that the deeper dataset structure may not improve the performance but makes large difference in risk over fitting.

Apart from different types of channel extraction for representation of multi perspective heartbeat, a one Dimensional layer in the basic module is filtered with different lengths. The ECG waves are fitted using different filters of different wavelength such as QRS complex and P wave. The max pooling layer is of different sizes which are followed by the 1D-convolution layer among the modules M11, M21, and M41. Different max pooling layers consists of outputs which are integrated with each other.