SMI-DSNet: Spatio-Temporal Matrix Imaging and Dual-Stream Networks for Enhanced Heart Sound Signal Classification

Heart sound signal analysis is a crucial tool for diagnosing cardiovascular diseases, where accurate classification is essential for clinical decision-making. However, current heart sound recognition methods lack sufficient interpretability and cannot fully utilize the information of time series data from sensors. To overcome these challenges, we propose the Spatio-Temporal Matrix Imaging and Dual-Stream Network (SMI-DSNet), a novel framework that integrates Spatio-Temporal matrix representations with a Dual-Stream deep learning architecture. The innovative Spatio-Temporal Matrix Imaging (SMI) technique transforms one-dimensional heart sound signals into two-dimensional matrix images, capturing key features that are clinically relevant for diagnosis and thereby enhancing the interpretability of heart sound data analysis. The Dual-Stream network integrates ResNet-18 for spatial feature extraction with BiGRU to capture long-range temporal dependencies in heart sound signals, thereby comprehensively harnessing the information encoded within the time series. Additionally, a cross-attention module is introduced to refine feature fusion by emphasizing pathological features based on physiological insights. Evaluated on public datasets, SMI-DSNet attains an exceptional classification accuracy of 99.5%, outperforming contemporary methods including CNN-BiLSTM and Vision Transformer. These results highlight the high precision of the model, demonstrating its potential for clinical application.