Heart Murmur Detection in Phonocardiographic Signals Using Breathing Noise Suppression

The automated heart murmur detection has become increasingly relevant to aid the medical diagnosis. The George B. Moody PhysioNet Challenge 2022 proposed a similar problem to encourage research in this field Our team's (AIMS-LAB) approach focuses on an accurate classification method using segmented S1, S2, systole, and diastole sections. A custom segmentation process was developed based on wavelet decomposition and Shannon entropy (WT-SE) for detecting the heart sounds, then later the systole and diastole regions could be determined. Along with this algorithm, a Hidden Semi-Markov Model (HSMM) was trained for comparing the results. To help segmentation a breathing noise suppression step was also developed. Multiple types of features were obtained for each segment, in time domain, frequency domain, and other features (e.g. Lyapunov exponent). Dimensionality reduction on the extracted feature-space was performed with two approaches, principal component analysis (PCA) and analysis of variance (ANOVA). For classification a support vector machine (SVM) and random forest (RF) was tested. Locally a 10-fold cross-validation was performed. During the official phase the best validation scores were 0.574 for the murmur and 9178 for the out-come task. We did not receive an official test score.