DropConnected neural network trained with diverse features for classifying heart sounds

Abstract

A fully-connected, two-hidden-layer neural network trained by error backpropagation, and regularized with DropConnect is used to classify heart sounds as normal or abnormal. The heart sounds are segmented using an open-source algorithm based on a hidden semi-Markov model. Features are extracted from the heart sounds using a wavelet transform, mel-frequency cepstral coefficients, inter-beat properties, and signal complexity. Features are normalized by subtracting by their means and dividing by their standard deviations across the whole training set. Any feature which is not significantly different between normal and abnormal recordings in the training data is removed, as are highly-correlated features. The dimensionality of the features vector is reduced by projecting it onto its first 70 principal components. A 10 fold cross-validation study gives a mean classification score of 84.1% with a variance of 2.9%. The final score on the test data was 85.2%.