Automatic segmentation and classification of cardiac cycles using deep learning and a wireless electronic stethoscope

We present an automatic heart-sound segmentation and classification system using state-of-the-art deep learning techniques and deep neural networks. This system is implemented for use with a wireless electronic stethoscope, known as Stethee
^®
. The proposed system automatically identifies the endpoints of temporal events associated with specific cardiac cycles, from audio recordings of heart-sounds captured by Stethee
^®
. The detected events are automatically classified as either a first heart sound (S1), a second heart sound (S2), systole, diastole or unwanted noise. We demonstrate that our proposed system is able to consistently achieve an accuracy of over 95%, across multiple permutations of training and test data. Finally, we carry out clinical trials against an echocardiogram device operated by a trained technician. We show that our proposed system is as accurate as Doppler echocardiography by a technician, in identifying the full cardiac cycle, systole and diastole durations for every analysed patient.