Sistema classificador de sons cardíacos baseado em redes neurais artificiais

Abstract

The World Health Organization (WHO) asserts that cardiovascular diseases are, nowadays, the major cause of mortality in the world. For its part, the entity also observes that the improvement in processes of prevention, diagnostic and treatment contribute to the decrease in lethality rates. Furthermore, inside the group of heart diseases, some have the possibility of diagnose through the identification of heart murmurs, that can be detected by means of an exam of cardiac auscultation – where the professional listens to the sounds originating in the heart, through a stethoscope, and performs the analysis of the patient's condition. Therefore, this work focuses on propounding a computer system that executes the classification of these heart murmurs in normal and abnormal. For that reason, the system uses database Physionet, in which the cardiac sounds, both pathological and not, are applied to train and validate classifiers, based in artificial neural networks. In this way, three types of classifiers are proposed, underpinned by MLP, LSTM and CNN, which are submitted to the audios of the database. The validation of these models, for its turn, will be carried out by evaluating the performance of the classifiers in terms of predicting heart sounds, based on the application of validation techniques suggested by the bibliographies. Consequently, by the end of the process, the objective is to obtain a robust and reliable classifier model, which is capable of performing heart sound predictions efficiently. After applying the methodologies to the classifier models, the CNN model performed best, with an accuracy of 0.8947, associated with a sensitivity of 0.8915 and a specificity of 0.8973.