Quantitative Dynamic Analysis of Phono Cardiogram with auto segmentation and alignment for the Detection of Heart conditions

Abstract

The evaluation of sounds produced by the beating heart and the blood flow provide a useful method for the diagnosis of cardiovascular disease (CVD). The cardiac auscultation using the classical stethoscope is a well-known diagnostic technique for identifying heart abnormalities. However, the examination requires a qualified cardiologist, who discerns the vibrational sound produced during the cardiac cycle’s contractions and valve closure for anomalies in the heart's pumping function. A phonocardiogram (PCG) signal records both the normal and pathological heart sounds, such as S3, S4, and murmurs, obtained during a stethoscope-assisted cardiac auscultation. The manual auscultation diagnosis is reported to be highly subjective and requires extensive experience. To counter the subjectivity and the high percentage of diagnostic errors, computer aided diagnosis (CAD) systems that detect the presence of abnormal heart sound conditions can be of paramount importance. In the case of computational analysis, either a statistical-based model or a dynamic-based model could be used. A dynamic model is more interpretable than a statistical model since dynamic models consider the morphology that represents biomedical signals in both time and frequency domains. The main aim of this work is to differentiate the heart conditions using parameter obtained by fitting the dynamic model to the experimental data. The dynamic model we used consisted of three coupled, ordinary differential equations (ODE). During the preparation of experimental data, the cardiac cycles were segmented at the correct positions considering the offset as well as onset of peaks and aligned using an algorithm called Icoshift. In our analysis, a total of 895 cardiac cycles were segmented; among which the count of normal beats containing only the S1 and S2 were 149; and abnormal cases containing S3, S4, or murmur were 746 beats. The dynamic model used is this work was modified from the reported literature. The parameters of the dynamic model was optimized by fitting it to the experimental data using Genetic algorithm. Besides, in the case of murmur, the model was used to fit the S1, and S2 segments and the fit data was subtracted from the experimental to recover the murmur signal. The comparison of the time domain correlation between the fit models and the experimental observations, were above 0.9 for all cases, validating the model’s accuracy.