Seismocardiogram-Based Classification of Cardiac Stenosis and Regurgitation Using Explainable Deep Learning Model

Effective early screening is crucial for preventing the progression of valvular heart diseases (VHD), such as stenosis and regurgitation. Deep learning (DL) techniques offer a powerful solution for capturing the complex attributes of VHD that are often difficult to identify with manually engineered feature learning methods. However, the black-box nature of DL models must be addressed for clinical acceptance. This study proposes an explainable DL framework designed to classify aortic stenosis and mitral regurgitation using seismocardiogram (SCG) signals. The architecture features four convolutional neural network (2D-CNN) layers with a residual connection to effectively capture and learn the intricate attributes of stenosis and regurgitation cases. Initially, 1D-SCG signals are segmented into artifact-free segments and transformed into SCG images by applying the normalized power of the continuous wavelet transform (CWT) coefficients to each segment. These SCG images are then input to the first CNN layer to learn low-level features, with subsequent CNN layers capturing more complex patterns. Our model demonstrates outstanding performance, achieving an aggregate accuracy of 98.02 %, a sensitivity of 97.06 %, and a specificity of 98.52 % when evaluated on a cohort of 16 patients with aortic stenosis, 13 with mitral regurgitation, and 29 healthy individuals. This method shows comparable performance to existing VHD detection approaches. Furthermore, Grad-CAM heatmaps provide visual interpretations of the model's decision-making process. This study underscores the effectiveness of SCG signals as a valuable tool for early valve disease screening.