Heart Murmur Detection using Supervised Machine Learning

Abstract

Murmurs are unusual heartbeat sounds that can be used to aid in diagnosing underlying health problems. Doctors often will manually perform heart auscultations, that is, attempt to hear these sounds using a stethoscope. This can be inaccurate as these murmurs are very subtle and can be muffled by background noises. Plus, it requires training, and the skill is easily lost if not practiced. It also requires having an appointment with a doctor, which is time consuming and sometimes inefficient. However, when successfully performed, it can provide valuable insights about the heart functionality of a patient. These murmurs can be either innocent which are safe, or abnormal. Heart murmurs are abnormal heartbeat sounds that can be shown in systole or diastole cardiac pathologies, yet manually diagnosing is inaccurate all the time. Hence by identifying murmurs we can classify them since they play a significant role in the diagnosis of a certain type of disease. Hence this project aims to present a system that enables to detect murmurs easily with efficient results and people can use it remotely without the need to visit their doctors. In this research, we proposed supervised machine learning specifically Mel Frequency Cepstral Coefficient to solve the previously mentioned problem statements, the main objectives are: To segment the systolic and diastolic phases of a heartbeat, to identify the heartbeat sounds as it is mixed with background sounds and to classify the abnormal heartbeat sounds with murmurs present from the normal ones. When trained and validated, the algorithm showed results of that train score 70.0%, test score 68.0% and validation 70.0%.