Detecting Aortic Stenosis Using Seismocardiography and Gryocardiography Combined with Convolutional Neural Networks

Aortic Stenosis (AS) is a heart valve disease characterized by the narrowing of the aortic valve opening. Currently AS is primarily diagnosed using echocardiography performed by a trained specialist. We aimed to evaluate the ability of non-invasive microelectromechanical system (MEMS) based seismocardiography (SCG) and gyrocardiography (GCG) sensors to detect AS in individual cardiac cycles in subjects by measuring the cardiac-induced vibrations produced by the mechanical activity of the heart. Data was collected from 20 AS subjects and 51 healthy subjects using a custom data logger capable of measuring SCG, GCG, and single-lead ECG. The captured SCG and GCG signals were segmented into individual cardiac cycles. A continuous wavelet transform was applied to produce time-frequency representations of each cardiac cycle. Each SCG and GCG axis of motion was then overlaid and fed as an input to a convolution neural network (CNN). Using leave-subject-out cross validation, the model produced specificity of 98.42%, sensitivity of 98.14%, and average accuracy of 98.36%.