Heart diseases are among the most important causes of mortality in the world, especially in industrial countries. Using heart sounds and the features extracted from them are among the non-aggressive diagnosis and prognosis methods for heart diseases. In this study, the time-scale, Cepstral, frequency, temporal and turbulence features are saved and extracted from the heart sounds, and then they are given to the multi layer perceptron neural network in order to be classified. Two methods, namely the UTA feature selection method and the genetic algorithm are separately performed for feature selection, and by introducing the effective features, it will be shown that in the best classification accuracies of 96 and 83 are achieved for the i-stethoscope and the digital stethoscope recorded heart sounds respectively. Totally when selecting the features using the UTA algorithm, a 4.25% increase has occurred on average in the classification accuracy for the i-stethoscope. Also, in the genetic algorithm, approximately 0.75% increase has occurred on average in the classification accuracy by selecting only 7 features. Manuscript profile