Page 120 - Control Theory in Biomedical Engineering

106 Control theory in biomedical engineering

assistance is therefore desirable by cardiologists for classifying cardiac
arrhythmias (Hijazi et al., 2016). According to the literature, the use of
DSS in cardiology has demonstrated their performance compared to manual
practice (Luz et al., 2016). However, developing, validating and implement-
ing such systems for healthcare today is a challenge for current research in
computer science, signal processing and medicine. Indeed, in order to
ensure the accuracy and the speed of the DSS results response, optimal
adjustment regarding the input hypothesis and machine-learning approaches
is required.
In this context, the main purpose of this chapter is to improve the accu-
racy of DSS based in cardiac arrhythmias classification. Indeed, several
e
machine-learning approaches are used for building DSS (Minchol et al.,
2019). The most common are neural networks (Vasilakos et al., 2016), fuzzy
logic controllers (FLC) (Ahmadi et al., 2018), support vectors machines
(Rajesh and Dhuli, 2017), recurrent neural networks (Singh et al., 2018),
k-nearest neighbors (Faziludeen and Praveen, 2016), genetic algorithms
(GA) (Li et al., 2017), decision trees (Kasar and Joshi, 2016), clustering
algorithms (Sayilgan et al., 2017), and many others (Mond ejar-Guerra
et al., 2019; Jambukia et al., 2015).
Recently, neural networks, which are inspired from the human neural
system, are mostly applied for classifying certain arrhythmias (Luz et al.,
2016). The most used are the multilayer perceptron (MLP) (Savalia and
Vahid, 2018), the radial basis function (RBF) (Kelwade and Salankar,
2016) and the probabilistic neural network (PNN) (Guti errez-Gnecchi
et al., 2017). The result of all the experiments done in (Lassoued and Ketata,
2018c) indicates that the RBF is the most accurate network with an accuracy
of 99.9%, while the MLP has fast testing response (0.096s) and the PNN has
the speediest training response (0.070s). However, it is concluded that there
is a significant relationship not only between the neural network’s perfor-
mance and its configured structure but also with the input feature vector.
So, several studies were conducted to evaluate neural networks with differ-
ent training algorithms (Lassoued and Ketata, 2018a) or number of hidden
neurons and layers (Lassoued and Ketata, 2017), or different datasets
(Rajamhoana et al., 2018) and many other factors (Li et al., 2017). Never-
theless, identifying the neural network structure, by using optimization
approaches, shows better performances (Lassoued and Ketata, 2018b),
(Arabasadi et al., 2017). Particularly, the MLP is mostly used in deep learn-
ing, which has advanced rapidly due to its structure in which both feature
extraction and classification stages are performed together (Bakator and