Abstract
Auscultation, the listening process for lung sound using acoustic stethoscope, is the first physical examination used to detect any disorder in heartbeat system. Unlike sophisticated tools, stethoscope is not beyond the reach of rural hospitals and clinics. However, the use of acoustic stethoscope needs specialized and well-experienced physicians. This is mainly due to limited sound amplification of the stethoscope to the extent that the human ears may fail to recognize the pathological sound, and hence, the diagnosis may be erroneously classified. Models that make use of screened digital, instead of acoustic, stethoscope, in which heart sound is digitized and stored, becomes one of the most popular techniques because it allows computer-aided software to perform automated analysis. In this paper, a complete algorithm for automatic heartbeat detection and disorder discrimination is presented. The technique takes the advantage of spectral analysis to separate the first and second heart sounds (S1 and S2) using a power threshold. The frame duration is dynamically estimated, according to duration of the sound to be analyzed (S1 or S2). As typical recordings of heart sounds are periodic with several cycles, two methods to combine MFCC estimates are proposed. Using 450 cardiac ausculatory with both pathological and normal heartbeats, the proposed methods were examined using a cross-validation strategy based on tenfold. More than 90%, at best, sensitivity and specificity, respectively, were obtained for the two methods using an artificial neural network classifier with multilayer perceptron. The solution takes the advantage of recent technology and digital advances, in which it is possible to connect the digital stethoscope to any digital device to conduct further analysis using computer-aided applications. The technique is practical as it can be available at different hospitals and clinics, including those in rural areas, with limited resources.