DropConnected neural networks trained on time-frequency and inter-beat features for classifying heart sounds

Automatic heart sound analysis has the potential to improve the diagnosis of valvular heart diseases in the primary care phase, as well as in countries where there is neither the expertise nor the equipment to perform echocardiograms. An algorithm has been trained, on the PhysioNet open-access heart sounds database, 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. Following this, the time-frequency behaviour of a single heartbeat is characterized by using a novel implementation of the continuous wavelet transform, mel-frequency cepstral coefficients, and certain complexity measures. These features help detect the presence of any murmurs. A number of other features are also extracted to characterise the inter-beat behaviour of the heart sounds, which helps to recognize diseases such as arrhythmia. The extracted features are normalized and their dimensionality is reduced using principal component analysis. They are then used as the input to a fully-connected, two-hidden-layer neural network, trained by error backpropagation, and regularized with DropConnect.

This algorithm achieved an accuracy of 85.2% on the test data, which placed third in the PhysioNet/Computing in Cardiology Challenge (first place scored 86.0%). However, this is unrealistic of real-world performance, as the test data contained a dataset (dataset-e) in which normal and abnormal heart sounds were recorded with different stethoscopes. A 10-fold cross-validation study on the training data (excluding dataset-e) gives a mean score of 74.8%, which is a more realistic estimate of accuracy. With dataset-e excluded from training, the algorithm scored only 58.1% on the test data.