for the classification while few other methods use the classical signal processing-based transforms to extract the

discriminating features. This paper proposes a novel integrated approach to classify the ECG signals for

arrhythmia detection. Unlike existing methods, it considers the multivariate time series nature of the input

along with the interrelationships among different ECG leads. The approach utilizes multivariate time series

features extracted using ROCKET (RandOM Convolutional KErnal Transform) and introduces new

connectivity-based features such as correlation and coherence for improved ECG signal classification. The

state-of-the-art classification performance of the proposed integrated model on the PTB-XL PhysioNet dataset

attested to the efficacy of the same.

Key-Words: - Arrhythmia detection, cardiovascular disease, disease diagnosis, electrocardiogram, multivariate

Cardiac arrhythmia, which refers to irregular heart

rhythms, is a prevalent concern in the field of

cardiology. It encompasses a spectrum of rhythm

abnormalities, some of which are harmless while

others could pose serious threats to health. As a

result, the prompt identification of arrhythmia holds

significant clinical significance and has the potential

to be life-saving, [1], [2]. In cardiology, the most

widely employed approach to identify cardiac

arrhythmia involves the use of the

introduces the possibility of human error and

observer bias, [2], [3]. To mitigate these challenges,

various methods have been proposed in the

literature to automatically identify cardiac

arrhythmia from ECG signals using the concepts of

authors utilized the discrete wavelet transform

(DWT) to extract features, combined with a support

vector machine (SVM) as the classifier, resulting in

an accuracy of approximately 95.9%. [6],

introduced a convolutional neural network model

for the purpose of categorizing 17 different cardiac

rhythms. Their reported accuracy stood at 91.3%,

derived from the analysis of 1,000 ECG fragments.

In their another work, the authors took a different

approach, utilizing both convolutional layers and

QRS peak interval, in conjunction with an SVM

classifier, resulting in a classification accuracy of

91.0%. Meanwhile, [9], introduced a novel deep-

learning methodology involving the transformation

of 1-D ECG signals into 2-D time–frequency

spectrograms using a superlet transform (SLT). This

innovative approach yielded an impressive overall

accuracy of 96.2%. [10], used an LSTM model to

identify 12 different heart rhythm categories. This

was achieved by analyzing a dataset consisting of

from ECG signals and was able to get an accuracy

of about 81.2%.

solve the arrhythmia detection problem, none of

these existing methods take into consideration the

multivariate time series nature of the input multi-

lead ECG signal and the interrelationship among

these leads. To circumvent these limitations of the

existing approaches, in this paper, we propose an

innovative approach to classify the ECG signals

or variance as representations. The ROCKET’s

convolutional kernels replace the need for manually

crafting these representations with a single

mechanism that can capture similar features

effectively. Furthermore, unlike the existing state-

of-the-art methods, here for the first time, we

2) Finance: It can be used to analyze the stock

3) Earthquake Warning Systems: Utilizing

5) Environment Monitoring and Fault Detection: It

time series classification, we next discuss ROCKET

(RandOM Convolutional KErnal Transform), one of

the best performing multivariate time series

4) Phase-lag value: Phase-lag value (PLV)

used a majority decision strategy, wherein each

subject was classified with a label of the majority

segments. For each sample, the 12-lead ECG signal

was divided into segments of 2-second length

(consisting of 1000 samples), and the predominant

class among all these segments was considered as

the class for the given sample. This approach helped

to mitigate potential fluctuations or anomalies

within shorter segments and ensured reliable

predictions.

Typical machine learning models often consider

every time series in a multivariate time series as an

independent entity, thus neglecting the relational

information amongst them, which can be exploited

for classification purpose. In particular, as far as the

multi-lead ECG signal is concerned, the neighboring

ECG leads can share meaningful relational

information, which can be quite useful for the

Fig. 2: Block diagram of the proposed connectivity-

based arrhythmia detection model

connectivity information from the ECG time series

using four widely used connectivity measures, viz.,

coherence, correlation, phase-lag index, and phase-

lag value. Each of these characteristics generates a

12x12 matrix, representing distinct qualities related

to each separate segment in the time series. Then we

classify into classes from this feature matrix, we

employed a convolutional neural network (CNN),

[19], as shown in Figure 2.

4 Experimental Results and

In this section, we first present the classification

performance of all three proposed models. Then we

compare their performance with the performance of

the existing arrhythmia detection methods. To

quantify the classification performance of the

detection model, we employed three performance

measures, viz., accuracy, sensitivity, and specificity

that are defined as:

where TP: True Positives represent cases where the

model accurately forecasts the positive class,

TN: True Negatives occur when the model

accurately predicts the negative class,

FP: False Positives happen when the model

inaccurately forecasts the positive class despite the

models to validate our results. By calculating the

average performance over several folds, the

fluctuation in the performance measurement is

minimized, enhancing its reliability and hence

providing a better estimate of how well a model is

likely to perform on unseen data.

multivariate time series-based model is concerned,

as the performance of the model depends on the

number of kernels in ROCKET and the ECG

(as shown in Figure 4).

used it as an input to the same CNN classifier. The

superior performance of the combined feature

matrix, as observed in Figure 5, verified our

hypothesis.

multivariate time series-based model and the

connectivity-based model, we next validate the

performance of our integrated arrhythmia detection

model which exploits both the sequential and

relational information of the ECG data.

Fig. 6: Classification performance of the proposed

arrhythmia detection models.

[5] Ch. Usha Kumari, A. Sampath Dakshina

[7] O. Yildirim, M. Talo, E. J. Ciaccio, R. S.

[8] K. Subramanian and N. K. Prakash,

[9] P. M. Tripathi, A. Kumar, M. Kumar and R.

[11] Q. Yao, R. Wang, X. Fan, J. Liu, and Y. Li,