Classification of Stenosis and Regurgitation from Seismocardiogram Signals using Convolutional Neural Network

Timely detection is critical in preventing the advancement of valvular heart diseases (VHD), such as regurgitation and stenosis. Deep learning frameworks present a robust approach for capturing the intricate features of VHD that are often challenging to discern with traditional manual feature extraction techniques. This research presents a convolutional deep learning framework designed to classify stenosis and regurgitation using seismocardiogram (SCG) signals. Our model incorporates four convolutional neural network (CNN) layers designed to extract and learn detailed characteristics of these conditions effectively. Initially, SCG time-series data are segmented to remove artifacts and then converted into SCG images using the normalized power of continuous wavelet transform (CWT) coefficients with analytic morse mother wavelets. The initial CNN layer processes these images to detect low-level features while the remaining layers extract more intricate patterns. Our approach achieves remarkable results, with an overall accuracy of 99.13%, sensitivity of 98.23%, and specificity of 99.42%, based on evaluations involving 17 patients with regurgitation, 18 with stenosis, and 29 healthy controls. This method outperforms current cardiac disease detection techniques, underscoring the effectiveness of SCG signals for early valve disease screening.