Power spectrum analysis for classification of heart sound recording

This study assessed the feasibility of using power spectrum analysis to compute additional features from heart sound recordings that can be used for normal / abnormal classification in the PhysioNet/Computing in Cardiology Challenge 2016. Our approach relies on the segmentation of the heart sound recording into the 4 distinct Fundamental Heart Sounds (FHS) states. Based on the FHS states, the heart sound recording is divided into multiple segments - each segment corresponds to 1 FHS state in 1 cardiac cycle. A fast Fourier transform is then performed on each of these segments. For each FHS state, we compute the ratio between the sum of the powers in the N highest peaks in the power spectrum to that of the entire power spectrum. The rationale is that the most dominant frequencies in the heart sound recording can potentially contain relevant information useful for classification. These additional features are subsequently combined with the features computed from the intervals of the 4 distinct FHS states. This enlarged set of features is used to train a feed forward neural network with 1 hidden layer for heart sound clarification. The scores of our neural network on a random subset of the test data are as follows: Sensitivity = 0.747; Specificity = 0.788; Overall = 0.767.