Cycle selection and neuro-voting system for classifying heart sound recordings

Aims: To address the PhysioNet/Computing in Cardiology (CinC) Challenge 2016 [1], this study aims to discuss a new method to classify PCG signals collected from a variety of clinical or nonclinical environments. Method: The raw PCG data is pre-processed by applying a band pass finite-duration impulse response filter (FIR) and discrete wavelet transform (DWT). For PCG delineation, first (s1) and second (s2) heart sounds are detected based on springer's algorithm. In this study, three analyses are used for feature extraction. The first one is time-frequency domain analysis in which the energy of nodes of wavelet packet transform is extracted beside the features already obtained from the time domain and perceptual domain analyses. Using fisher discriminant analysis (FDA) for feature reduction, a new space to better represent features is obtained. For heart sounds classification, a combination of three classifiers of artificial neural network (ANN) are implemented. To accomplish a precise classification, the output of all classifiers are employed in a fuzzy-logic algorithm (voting system). Results: By evaluation of the proposed method by training set the sensibility of 0.96, specificity of 0.97 and total score of 0.98 were achieved. Also, evaluation on the hidden test set resulted in the sensibility, specificity and modified accuracy of 0.811, 0.872 and 0.841, respectively.