International Journal of Scientific and Technological Research www.iiste.org ISSN 2422-8702 (Online), DOI: 10.7176/JSTR/5-12-15 Vol.5, No.12, 2019 138 | Page www.iiste.org Using Neural Networks for Heart Sounds Classification: A Review Omer Deperlioglu Department of Computer Programming, University of Afyon Kocatepe, Afyonkarahisar, Turkey E-mail: deperlioglu@aku.edu.tr Abstract The high number of sudden deaths from heart diseases today makes early diagnosis of heart diseases very important in order to prevent these deaths. The first method that doctors use to diagnose heart disease is to listen to heart sounds caused by mechanical movements of the heart. This procedure is commonly referred to as auscultation of the heart. Over time, the irregularity of S1 and S2 sounds, which are components of heart sounds, may show some disturbances in heart function. For example, inadequate heart valves or murmurs. Early diagnosis of heart diseases is of great importance. However, not all doctors have the ability to interpret every heart sound and make an early diagnosis. Of course, the fact that their jobs are very busy and too many patients affect this. The status of each patient may be different during diagnosis and treatment. Assistive systems have been developed to enable doctors to diagnose patients with very different symptoms. These systems generally take diagnostics and evaluate all the data and suggest a diagnosis for the patient. In other words, these systems are decision-making mechanisms that classify all data. In the classification studies, researchers used different classification techniques in order to increase the classification success. Because the success of the classification means an increase in the diagnostic ability. For this purpose, many different techniques and mixed methods are used. In this study, a general evaluation of my classification studies of heart sounds with different neural networks such as Artificial Neural Networks (ANN), Convolutional Neural Network (CNN) and Autoencoder Neural Networks (AEN) was performed. PASCAL B-Training and Physiobank-PhysioNet A-Training data sets were used for the classification of heart sounds. The efficiency of the methods were compared by considering the common evaluation criteria in the studies examined. Keywords: Heart sounds classification, Re-sampled signal energy, Autoencoder neural networks, Artificial Neural Network, Convolutional Neural Network. DOI: 10.7176/JSTR/5-12-15 1. Introduction According to the World Health Organization (WHO), approximately 30% of all deaths in the world are caused by heart diseases. Therefore, any method to detect symptoms of heart disease can have a significant impact on world health. Numerous cases of medical data, as well as very different diseases and symptoms, make it difficult to examine and investigate the data. Early diagnosis of heart diseases is of great importance. However, not all doctors have the ability to interpret every heart sound and make an early diagnosis. Of course, the fact that their jobs are very busy and too many patients affect this. The status of each patient may be different during diagnosis and treatment. Assistive systems have been developed to enable doctors to diagnose patients with very different symptoms. These systems generally take diagnostics and evaluate all the data and suggest a diagnosis for the patient. In other words, these systems are decision-making mechanisms that classify all data. In the classification studies, researchers used different classification techniques in order to increase the classification success. Because the success of the classification means an increase in the diagnostic ability. For this purpose, many different techniques and mixed methods are used (WHO, 2017; Deperlioglu, 2019a). The majority of classification procedures are performed to achieve better classification results. Some examples of the work done for the classification of heart sounds other than my own studies are given below. Shankar and Sangeetha In their study using artificial neural networks for classification, S1 and S2 segmented heart sounds using discrete wavelet transform and Shannon energy envelope. They tried to