Exploring Feature Selection and Classification Algorithms

For Cardiac Arrhythmia Disease Prediction

RAVINDER AHUJA1, SC SHARMA2

Abstract: Cardiac Arrhythmia is the disease in which heartbeats abnormally due to which death of a

person may occur if not diagnosed on time. Timely and accurate detection of cardiac arrhythmia can save

the life of the patient. In this study fourteen classification algorithms and six feature selection algorithms

Selection, multiclass Classification.

Received: June 14, 2021. Revised: June 10, 2022. Accepted: July 8, 2022. Published: August 5, 2022.

1. Introduction

Heart disease have affected the significant amount of the

population, not only in India but all over the world. Any

sort of irregular behavior in the heart may be life-

threatening. A tool can be used for diagnosing the proper

activity of heart, known as ElectroCardioGram (ECG),

which produces a graph for electrical pulses [1]. Specific

parameters are taken into account, for the proper

to human life and may lead to a cardiac arrest. Therefore,

accurate detection and classification of arrhythmia are

essential. These conditions may cause the heart to beat

fast or slow, skipping some beats. Because of these types

of behaviors, ECG will form different graphical patterns,

and therefore, arrhythmia can be easily detected [3].

There are generally two different categories, one which

causes the heart to beat too slowly, usually below 60 rpm

known as Bradycardia, and another one causes the heart

to beat at 100 rpm, known as Tachycardia [4]. There are

difficulties for a doctor to look at those and find

irregularities [5]. Therefore, Machine Learning can be

used

for the automation of arrhythmia diagnosis, and it will be

quite helpful. In this paper, six feature selection

techniques are applied to reduce the dimension and

further fourteen classification algorithms, and their

ensemble has been applied to classify into one of the

sixteen classes. The major points of the work done in this

study is as follows:

is applied with different values of k (2, 5, and 10). Best

results are reported at K=10.

The rest of the paper is divided into the following

sections: section 2 contains related work, section 3

contains dataset description, preprocessing techniques

used, feature selection techniques used, and classification

algorithms, and methodology used, section 5 contains

experimental results, and section 6 contains conclusion.

2. Related Work

automation of ECG analysis and recording the patient's

health status was reported by [7]. Some of these methods

that have helped in the development of ECG analysis are

algorithm is applied for the classification of arrhythmia

[11]. In paper [12], authors have used a correlation-based

feature selection technique to select essential features

from the UCI repository dataset. The neural network

algorithm is applied along with the Levenberg-Marquardt

method for the classification of arrhythmia. The random

forest classification algorithm is implemented with

resampling technique is used to classify cardiac

arrhythmia [13]. In paper [14], firstly, various

preprocessing techniques are applied first. Various feature

Analysis) feature selection technique is used to extract

essential features, and further SVM classification

algorithm is applied to classify arrhythmia. In paper [19],

authors have applied neural networks on the dataset for

prediction of cardiac arrhythmia and achieved an accuracy

of 76.67%. In paper [20], authors have applied the Naïve

Bayes classification algorithm with the train-test split

ratio of 70-30 and achieved an accuracy of 70.50%. In

paper [21], authors have applied feature selection

techniques in two steps: the wrapper method part and the

selected. Further, modular Neural Network with three

layers was used for classification and achieved an

accuracy of 78.89% with a train-test split ratio of 90-10.

3. Materials and Methods

The methodology applied has different components which

are as follows:

the UCI machine learning repository [23]. The dataset

consists of 452 rows corresponding to each patient. Two

unlabeled patient. Two hundred forty-five rows are

corresponding to class 1 label, remaining 185 classes

missing values and outliers. The missing values are filled

with mean. The outliers are detected using the

interquartile range and handled with logarithm. A

standard scalar is used to normalize the data.

3.3. Feature Selection Methods: The dataset

Figure 1: The overview of the methodology use

features is set up manually as 55, so it will return 55

features. It is observed from the results that random forest

classifier has produced better results with an accuracy of

76.87%, recall of 78.29%, precision of 78.29%, and F1-

score of 78.29%. Next to the random forest is the

maximum voting classifier giving an accuracy of 75.27%,

precision 77.21%, recall 77.21%, and f1-score 77.21%.

The next to maximum voting is stacking, giving an

accuracy of 70.83, precision 75.82%, recall 75.82%, and

f1-score 75.82%. The light GBM classifier has produced

observed from the results that random forest classifier and

maximum voting has produced better results among all

the classification algorithms and giving an accuracy of

74.22%, recall of 77.70%, precision of 77.70%, and F1-

score of 77.70%. The next best performing classifier is

stacking, giving an accuracy of 72.36%, precision

75.39%, recall 75.39%, and f1-score 75.39%. The worst

performance is given by SVM classifier, giving an

accuracy 55.34%, precision 58.24%, recall 58.24%, and

f1-score 58.24%.

0.25 is used through which 221 features were dropped. It

is observed from the results that random forest classifier

has produced better results with an accuracy of 80.22%,

precision of 79.12%, recall of 79.12%, and F1-score of

79.12%. Next to the random forest is maximum voting

and XGBoost classifier giving an accuracy of 72.83%,

precision 76.92%, recall 76.92%, and f1-score 76.92%.

The KNN classifier has produced poor results among all

with an accuracy of 58.05%, precision 60.73%, recall

60.73%, and f1-score 60.73%.

66.63%, and F1-score of 66.63%. Random forest is a

collection of decision tree methods. Each decision tree

constructs a classifier based on a random data sample, and

several classifiers are integrated to generate a single

classifier known as random forest.

Comparison of different approaches reported in literature

classifier since it produces its result by combining the

outputs of many decision trees (created by dividing the

training data into subsets). Furthermore, the random forest

feature extraction technique assisted us in selecting

features from a large category with high weightage in

categorising the cardiac arrhythmia. Figure 2 gives a

comparison of six different feature selection techniques

used in our study. The table provides the performance

parameters reported in each of the feature selection

techniques. It has been observed that the random forest

[27] Koller, D., &Sahami, M. (1996). Toward optimal feature

Applications (pp. 106-115).Springer, Berlin, Heidelberg.

[30] Schölkopf, B. (2005). Introduction to Kernel Methods. The

Analysis of Patterns, Erice, Italy.

convex distance nearest neighbor algorithms.In Advances

in neural information processing systems (pp. 985-992).

regression by randomForest. R news, 2(3), 18-22.

algorithm for text categorization.In Proceedings 2001

decision tree learning algorithm. In icml (Vol. 99, pp. 124-

G. C. (1990). Generalized averaging method for power-

conversion circuits. Technical report (No. AD-A-

machine learning.International workshop on multiple

International Journal of Information Sciences and

Techniques (IJIST), 2.

guide to the gbm package. Update, 1(1), 2007.

algorithm. Geophysical Prospecting, 27(1), 16-28.

[42] Rokach, L. (2010). Ensemble-based classifiers. Artificial

Intelligence Review, 33(1-2), 1-39.

majority voting. Information fusion, 6(1), 63-81.

modeling of ECG signal. Medical and Biological

Engineering and Computing, 34(2), 171-174.

invariants. Computational and mathematical methods in

medicine, 2018.