Page 92 - The-5th-MCAIT2021-eProceeding

3.1. Machine Learning

According to the literature, SVM and CNN are widely used in CVDs diagnosis using ECGs data. Both SVM
and CNN are considered powerful algorithms, and they achieved very high accuracy rates compared with other
methods (Ahmad et al., 2014). The main reason for SVM significant performance is the feature dimensionality
independency which makes it immune from the “curse of dimensionality”. SVM works well with a clear margin
of separation between classes. The hyperplane is affected by only the support vectors. Thus, outliers have less
impact, which makes it particularly efficient for classifying complex but small or medium-sized datasets (Li,
Bhanu, & Krawiec).
Furthermore, different SVM classifiers can be constructed using different kernels (polynomial, RBF, linear)
to solve leaner or non-leaner problems. However, selecting an appropriate kernel for a. given problem is still
under research. Besides, SVM does not work perfectly in a noisy or large dataset, and it takes a long time in
training the model (Bisong, 2019). In contrast, CNN can work perfectly in a noisy and large dataset because of
its capability of finding optimal temporal features better than other machine learning algorithms.
CNN can learn complex features by preserving the spatial relationship between the feature (Bisong, 2019).
Reducing the number of weights needed for training the network is another significant advantage of CNN.
However, both SVM and CNN are High computational costs and time and memory consuming (Alizadehsani,
Roshanzamir, et al., 2019).
A small number of databases are available in the public domain, which explicitly studied in most previous
works. Further, most of these public databases have a small sample size and unbalanced data. MIT-BIH,
ZAlizadeh Sani, Cleveland, and Hungarian datasets are the most popular public dataset. The main reason
traditional machine learning algorithms are significantly well performed is that most of the sample size of these
public domain datasets, accept MIT-BIH, is between 303 and below with few features. As seen from the
literature, a feature selection method has been used in most structured data works, and it has a significant impact
on model performance.
In signal data, a combination of the Fantasia dataset and St.-Petersburg dataset are widely used. The ECG
signals were retrieved from Fantasia (for Normal) and St.-Petersburg Institute (for CAD). However, most
datasets in this field are either very small in size or inaccessible to the public. One reason traditional ML has
worked sufficiently well in previous years in using ECGs to diagnose cardiology is that experts are carefully
designed feature extraction methods. Also, handcrafted features such as statistical measures from the ECG beats
and the RR interval positively influence traditional ML’s performance. Recently, using DL to diagnose CVDs
from ECG signals has gained much interest due to its simplicity and reduced dimensionality compared to
imaging data (Strodthoff, Wagner, Schaeffter, & Samek, 2020).

4. Summary

Signal data and especially ECG signals have gained a lot of interest in the research community with respect
to the other data type. ECG is a widely available and great tool used to diagnose CVDs with high efficiency
from a medical aspect. However, it is noisy sensitive and can be misleadingly interpreted by doctors. ML and
DL incredibly proved their ability to insight hiding information and provide simplicity in handling such
problems. CNN algorithms particularly have achieved high efficiency and outstanding performance in
diagnosing CVDs using ECG signals. This survey introduced the ECG structure and components, and the
diseases that can aid to diagnose. The related state-of-art works used to detect CVDs using ECG signals is
discussed as well as the widely used ML algorithms, feature selections, and datasets.

Acknowledgment

The authors acknowledge the support of this research by the Yayasan Universiti Teknologi PETRONAS
Fundamental Research Grant (YUTP-FRG) under Grant 015LC0-244

E- Proceedings of The 5th International Multi-Conference on Artificial Intelligence Technology (MCAIT 2021) [79]
Artificial Intelligence in the 4th Industrial Revolution