Abnormality Detection Method from Heartbeat Sound by Machine Learning

Heart disease is currently the second leading cause of death in Japan, and approximately 80% of patients with heart disease are aged 65 years or older. In addition, in the diagnosis of cardiac disease, highly accurate diagnosis is difficult in depopulated and aging areas due not only to the lack of medical equipment but also to the shortage of medical specialists. Therefore, the ultimate goal of this study is to create a system that can automatically diagnose patients with cardiac disease using machine learning from heartbeat sound data transmitted by the patient. As a preliminary step, to investigate effective features for discriminating normal and abnormal heart sounds from heart sounds collected using a digital stethoscope and iStethoscope Pro, and to study the effects of discriminator identification and noise removal on the identification results. The outline of the proposed abnormality detection method is to perform noise processing using spectral gating and a zero-phase filter on the loaded auscultation data. Next, using openSMILE, statistics are calculated from three feature quantities, MFCC (0-14th order), logMelFreqband (0-7th order), and lspFreq (0-7th order), and PCMloudness, which is related to the “volume” of the PCM data. Finally, using WEKA, which is used in machine learning and data mining, a 10-fold cross-validation using SVM and Random Forest (RF) is used to evaluate whether the heart sounds are abnormal or normal. In this study, SVM was shown to be more effective than RF as a discriminative model and spectral gating was more effective than zero-phase filter in noise processing.