Arrhythmia is ubiquitous worldwide and there is still a major population at risk. The ECG signals are highly efficient and used as a gold-standard to detect the presence of arrhythmia and critical conditions such as cardiac arrest. So, detecting arrhythmia using ECG signals is challenging for researchers. In this era of automation, deep learning has made strides in various fields such as computer vision, language processing, and signal processing in developing state-of-the-art models on large scale databases. Further, deep learning has advanced in bio-medical imaging and bio-medical signal processing. Hence, it is aimed to develop a diagnostic model by extracting features from ECG using DT-CWT and processing them with help of the proposed neural architecture.

It is observed that the signals recorded by ECG are a combination of PQRS waves and these waves detect the heart functionality by identifying various characteristics. Certain features are extracted from the signal by pre-processing it with various transformation techniques in which continuous wavelet transform and discrete wavelet transform are frequently used. Sequentially these extracted features are then processed using various learning algorithms such as k-nearest neighbors (k-NN), Support vector machines (SVM), Multi-layer-perceptron (MLP), Singular value decomposition (SVD), etc. The pre-processing techniques are implied for extracting the features from ECG signal widely used Discrete wavelet transformation (DWT), which is sensitive to shift-invariance and downgrades the quality of the signal while reconstructing the decomposed signal. In the subsequent processing step, most of the methods include generic machine learning algorithms (classification or clustering algorithms) or an MLP which does not capture temporal invariances and eventually harms the performance by reducing the generalization. Hence the two key steps to provide a diagnostic model are, (a) an appropriate pre-processing of the signal (DT-CWT) (b) a processing step to prognosticate the disease (neural attention). To overcome the drawbacks of the existing research, the proposed diagnostic framework implies DT-CWT and neural attention mechanism to provide a significant solution.

The impact of MIT-BIH data was shown by George et al. [1], which acted as a catalyst and gave rise to numerous works that provide insights on automated devices to diagnose arrhythmia. Markos et al. [2] used time-domain analysis to extricate features and then arranged them in distinct combinations, which are utilized as input for neural networks. Sixty-three different types of neural networks were formed. The output of these networks was deployed to a decision tree to diagnose arrhythmia. Karimifard [3] worked on modelling of signals, who later used a Hermitian basis function to get a feature vector and sent to a k-nearest neighbor classifier to classify seven types of arrhythmia, which [1–4] obtained sensitivity and specificity of 99.0% and 99.84% respectively. He also concluded that the size of the feature vector affected the training time of the model. Mohammadzadeh et al. [4] took features from the signal by linear and non-linear methods, which were reduced by Gaussian discriminant analysis (GDA) and used an SVM to recognize six classes of arrhythmia with a sensitivity of 95.7% and specificity of 99.40%. Chi et al. [5] focused on quick prediction of the disease by pulling out PQRST features from the ECG signal and later used Linear Discriminant Analysis for grouping five different classes of arrhythmia and achieved an accuracy of 96.23%.

A unique use of the kernel Adatron algorithm was combined with SVM by Majid et al. [6]. He explained the drawbacks of a multi-layered-perceptron (MLP) and differentiated the training and testing time of these two methods. Hamid et al. [7] performed Complex wavelet transformation (CWT), Discrete wavelet transformation (DWT), Discrete cosine transformation (DCT) feature extraction methods separately on the signal and formed four different structures using MLP, then the other four using SVM and deduced the efficient use of a feature extraction method by the training time of the model [5–10]. Oscar [8] preprocessed the signals by QRS extraction method and used fuzzy KNN, MLP with backpropagation, and MLP with scaled conjugate gradient backpropagation (GBP) to get the output matrix. Later these three matrices were combined and sent to the fuzzy inference system to get the result. This achieved an accuracy of 98%. Roland et al. [9] presented an artificial neural network that took signals that are preprocessed by Fast Fourier Transformation (FFT) as an input and then categorized five classes of arrhythmia.

The importance of PQRST wave properties was also discussed here. Yeh et al. [10] proposed a novel preprocessing method along with Cluster analysis (CA) to classify 5 distinct classes and attained a total classification accuracy (TCA) of 94.30%. Stefan [11] developed an android application for real-time detection of arrhythmia by Decision Trees (DT). This model clocked a sensitivity of 89.5% and specificity of 80.6%. Elgendi [12] introduced the application of the moving averages method on ECG signals for detection of P and T waves by addressing four sources of noise which altered the quality of the signal and obtained a sensitivity of 98.05% and specificity of 98.86%. Manu et al. [13] extracted features using DTCWT and merged another four features (AC power, Kurtosis, Skewness, and timing information). This feature set was passed into an MLP and got an accuracy of 94.64% and a sensitivity of 94.6%. Ahmet et al. [14] showed a comparison between the performance of bagged decision trees and a single decision tree with the input of nine features which was taken from ECG signal by applying Low Pass Filter, High Pass Filter, form factor (FF) computing, FF ratio to previous one (FFR), RR ration to the previous RR ration (RRR), RR difference from mean RR value (RRM), skewness, linear predictive coding (LPC) and the cumulated ensemble method outperformed a single decision tree with an accuracy of 99.15%. Mehrdad [15] improved the signal quality by un-decimated wavelet transformation (UWT) and then proposed a method that combined Negatively Correlated Learning (NCL) and Mixture of Experts (ME) which is known to provide an excellent recognition rate. The model is used to group premature ventricular contraction (PVC) arrhythmia and Normal heartbeat classes and achieved accuracy, sensitivity, the specificity of 96.02%, 92.27%, and 93.72% respectively. Ping [16] proposed an adaptive feature extraction method based on wavelet transformation and a modified voting mechanism consists of K-means clustering, one against one SVM to enhance the recognition rate and got an accuracy of 89.2%. Joachim [17] worked on the categorization of poor and good signals. An alarm is set off when the parameters are not within a given scale. QRS extraction method along with SVM was used to reach this objective. Patricia [18] presented a Learning Vector Quantization (LVQ) algorithm with SVM to classify arrhythmia. However, the comparisons were made with simulated data. A total of 15 classes were grouped with three different architectures and the best architecture got an accuracy of 99.16%. Ali et al. [19] performed a diagnosis of arrhythmia with the help of Alex net and prior QRS detection was done. The signal was converted to a 256 × 256 sized image and then passed into the network. The recognition rate and accuracy are 98.5% and 92%. Joy [20] implemented DCT transformation of waves and used Probabilistic Neural Network (PNN) for efficient detection of disease [11]. Vasileios [21] showed false beat detection effectively by detecting QRS peaks then filtering false beats using SVM and concluded that QRS peaks are very important for the detection. Rashid et al. [22] proposed a new method that showed promising results by using Gaussian mixture modeling (GMM) with expectation maximization (EM), Combined with statistical and morphological features. Accuracy for class-oriented is 99.6% and for subject-oriented is 96.15%. Serkan et al. [23] also made an android application using 1-D convolution to classify supraventricular ectopic beat (SVEB) and ventricular ectopic beat (VEB). FFT with DCT was used in feature extraction and obtained an accuracy of 99.0%, 97.2% for each class.

From the previous research works, it is observed that (pre-processing part) many methods do not capture temporal relationship among the data. If they capture temporal dependencies, they do not persist with long term dependencies. So, if long term dependencies are provided there are no sequential patterns, which provide attention to the network determining the importance factor. So, these loops are overhauled with the use of attention embedded neural architecture by capturing long term temporal dependencies. Further, some loops are addressed in the pre-processing section and they are overridden by utilising DT-CWT and are mentioned in the successive section.

In this research, the contributions to the body of the knowledge are mentioned as,

Dual-Tree Complex Wavelet Transform (DT-CWT) method is implied to overhaul shift-invariance and aids the signal reconstruction to extract significant features. Further, a small set of features are extracted using the Pan-Tomkins algorithm and are adjoined with the features extracted from DT-CWT.

As mentioned, the two important steps are involved to complete the proposed automated system. This section aims to give a clear understanding of mathematics related to these two steps and explains the unique capabilities.

As mentioned above, to leverage the model performance, training and testing samples are split into multiple variants ranging from 10% to 50%. The complete analysis is carried out on various standard classification metrics and to study the proposed model behaviour, an accuracy score is chosen as the gold standard. Similarly, to study class wise performance precision, f-1 score recall is utilized. MSE is used to assess the predictability of the model which depicts the error attained due to imperfect predictions. Finally, AUC-ROC curves are generated to assess the diagnostic performance of the proposed model [41].

As AUC-ROC curves are sample invariant in nature as they are insensitive to the alterations implied in the class distributions. These curves are plotted class-wise to interpret the performance of the model at each class level. As a note, AUC-ROC visualizations can be obliged as they decouple the performance of the classifier from skewness in classes and error costs presented in Fig. 8. The proposed model is trained with two variant batches of 32 and 64 respectively. All the above-mentioned classification metrics are evaluated for all the test variants with two different batch sets. Large batches are acquired during training neural networks to minimize the generalization gap [42]. Hence, a large set of batches are considered with sizes of 32 and 64. (illustrated in the Tab. 1). In the feature extraction step, the Pan Tomkins method is employed to extract QRS points of the signal which play an important role in determining r-peak which helps to detect heartbeats. A large set of features are drawn out by using DT-CWT. This transformation is shift-invariant and provides PR. Most of the signal transformation methods lack these properties which can cause imperfect prediction and increase the chance of misclassification.

The proposed network used adam [43] as an optimizer with a learning rate of 10−3. The neural network uses categorical cross-entropy as the objective function for stochastic optimization of neural network with backpropagation.

(20) Loss=−∑i=1kyilog⁡(yi^)

Generally, RNNs understand the temporal dependencies but they lack understanding in long term dependencies where LSTM overcomes the problems of RNN by understating long-term temporal relationships in the data. But LSTMs are computationally expensive and do have the problem of gradient vanishing with increasing units to a greater extent. Whereas GRU contains fewer gates compared to that of LSTM networks and overcomes the problems of LSTMs. GRUs are computationally faster compared to RNN’s and LSTMs. As most of the research focuses on designing a neural architecture utilizing these units with changing the number of time steps, units and stacking pattern. To leverage predictive capability neural attention is implied by undersetting the temporal pattern extracted from the signal. So, by providing neural attention, the minute redundancies and noise captured during feature extraction can be regulated to a greater extent. This provides greater performance compared to that of remaining neural networks. Various previous work is studied and curated, and our method outperforms the existing literature, and the depicted results are tabulated (illustrated in Tab. 2).

As a note, the research was conducted to study arrhythmia without using signal patterns i.e., carried out by classifying variant attributes involved in predicting cardiovascular diseases [44] and also carried by implying PPG signals [45]. To see the future perspective of the proposed work, it can be figured out traditionally, the current deep learning applications have considered existing distance functions in the research literature for similarity computations but did not try to fit in new functions for similarity computations [46–50]. There is a possibility to devise threshold and similarity functions to suit deep learning applications [51–55]. For instance, recent research contributions propose various similarity and threshold functions for temporal pattern mining which can be redesigned to suit deep learning applications [56–60].

In this research, a novel attention-based neural architecture is built to vanquish the loops of existing methods for classifying ECG signals. It can be stated that the proposed model is sample invariant as it has minute error variation when test samples increased five times. AUC-ROC plots are illustrated to provide a vivid understanding of the performance of the proposed diagnostic model. In a worst-case scenario, the model provides a micro averaged AUC of 0.9904. Even with numerous advantages, it is seen that the proposed model can consume high memory while embedding the model into a real-world application. The training procedure adapted is tested on two batches instead of a dynamic sampling is preferred to improve performance. In future, it is aimed to provide a salient model by acquiring humongous data with less computational capability and higher performance.